• 专利附录
  • 专利说明
  • 权利要求
  • 类似技术
  • 同族专利
  • 引用文献
  • 法律状态
  • 优先权
    • 专利摘要:
    [Object] The invention relates to an inexpensive and secure moving object detection apparatus and a moving object detection method which enable precise detection of a moving object from a moving image of a monocular camera. at a high speed. [Means for Solving Problems] A representative configuration of the mobile object detection apparatus according to the present invention is provided with a horizon line detection unit 124 which detects a horizon line in a frame image , a rim image generation unit 122 which generates a rim image from a frame image, and a moving object detection unit 118 which defines a detection box on a moving object, and the ledge image generation unit extracts a ledge image under a horizon line detected by the horizon line detection unit, and the moving object detection unit generates a foreground by combining the difference between the edge image under the horizon line and a background image of the edge image with the difference between a grayscale image and a background image of the grayscale image. Figure 4
    公开号:FR3085219A1
    申请号:FR1909276
    申请日:2019-08-19
    公开日:2020-02-28
    发明作者:Takuya Akashi；Uuganbayar Ganbold；Hiroyuki Tomita
    申请人:Inc National Univ Iwate Univ；Suzuki Motor Co Ltd；
    IPC主号:
    专利说明:

    Description
    Title of the invention: MOVING OBJECT DETECTION APPARATUS AND MOVING OBJECT DETECTION METHOD
    Technical Field [0001] The present invention relates to a moving object detection apparatus and a moving object detection method which allow precise and high speed detection of a moving object from a moving image of a monocular camera.
    PRIOR ART Currently, research and development of a large number of preventive safety technologies aimed at preventing the occurrence of traffic accidents is progressing. In this regard, a need for prevention of frontal traffic accidents between an automobile and a pedestrian, a bicycle, another automobile, etc. at intersections, etc. In systems designed to meet such needs, it is conceivable that an on-board sensor or an on-board device is used to detect a moving object.
    [0003] As object recognition technologies which use an on-board sensor, systems exist in which an auxiliary device such as a laser distance sensor is used. However, since laser devices are expensive, and an intense laser should not be pointed at a human body, there is a problem that it is difficult to increase its power.
    [0004] When using an on-board camera, it is conceivable to use a background difference method. The background difference method is a reliable technique that allows processing at high speed, and uses changes within a dynamic scene. However, if an object has a flat texture, or the color of part of an object is similar to the background color, in some cases only part of the object can be detected.
    Document Patent 1 proposes an image processing apparatus which, if there is a plurality of candidates for the region to be detected (a front part and a foot part), specifies a size, in a captured image , to be occupied by a region to be detected, and, if another candidate is included in the size range, extracts a region obtained by coupling the candidates and the other candidate as a candidate to be a new region to be detected.
    Technical problem [0006] However, in the technology of Document Patent 1, a size to be occupied is specified on the basis of the positions of the candidates for a region to be detected, in a captured image. In this case, it is necessary to first determine the distance to the region to be detected. In Document Patent 1, an image capture device is attached and installed to monitor an intersection (paragraph 0038, Fig. 1). Due to this restrictive condition, the distance can be measured from a position in an image, but, in the case of an image from an on-board camera, it is difficult to measure the distance.
    [0007] In addition, in Document Patent 1, it is necessary to determine the type of region to be detected (whether the subject is a person or not), and to specify a size to occupy. Such a determination is very unreliable. Furthermore, in Document Patent 1, an infrared ray camera is used, and thus a part in which a skin is exposed is detected, but the unreliability increases depending on the fact that the surface of the exposed skin differs significantly according to the orientation of the person, his outfit and his hairstyle.
    In addition, when using an inexpensive monocular camera instead of an infrared camera or when using a hypergone lens or a wide-angle lens in order to take image of a wider range, noise may be higher than noise when using an infrared camera.
    In this regard, the applicants propose, in Japanese patent application No. 2017-178431, a mobile object detection device and a method of detecting a mobile object which are inexpensive and safe, and allow a precise and high-speed detection of a moving object from a moving image from a monocular camera. Furthermore, the present invention aims to provide a mobile object detection apparatus and a mobile object detection method further improving Japanese patent application No. 2017-178431, and in which the detection rate is improved. Technical solution In order to solve the problems described above, a representative configuration of a mobile object detection apparatus according to the present invention comprises an input unit for introducing a mobile image, a data acquisition unit frame which continuously extracts a plurality of frame images from a moving image, a horizon line detection unit which detects a horizon line in a frame image, an image generation unit ledges which generates an edge image from a frame image, and a moving object detection unit which defines a detection box on a moving object, and the edge image generating unit extracts an image of ledges below the horizon line detected by the horizon line detection unit, and extracts a difference between the image of ledges below the horizon line and a background image of the image of ledges using a different process background image, and the moving object detection unit converts a raster image to a grayscale image and extracts a difference between the grayscale image and a background image of the grayscale image gray using the background difference method, and generates a foreground by combining the difference between the edge image under the horizon line and the background image of the edge image with the difference between grayscale image and the background image of the grayscale image.
    According to the above configuration, it is possible to accurately detect the position in the image of a part in which the movable object is connected to the ground. So even if the object has a flat texture or is moving slowly, the whole object can be detected as the foreground, and the detection rate can be improved. Therefore, it is possible to accurately detect a moving object from a moving image of a monocular camera at high speed, and security can be improved at low cost.
    The mobile object detection unit preferably combines the difference between the image of ledges below the horizon line and the background image of the edge image with the difference between the image in levels of grayscale and the background image of the grayscale image, if the difference between the edge image under the horizon line and the background image of the edge image contains little noise .
    The rim image generation unit preferably extracts a rim image which is connected to the horizon line detected by the horizon line detection unit, and the unit for detecting mobile object preferably filters the foreground using the edge image which is connected to the horizon line. According to the above configuration, the boundary between a building and a road can be removed as noise.
    When the moving object detection apparatus moves, the moving object detection unit preferably filters the foreground using the image of edges which is connected to the horizon line. As the moving object detection apparatus moves, the image positions of a road and a building change, and therefore the noise caused by the background difference process increases. It is therefore possible to reduce the noise significantly by carrying out the filtering described above, when the moving object detection apparatus is moving.
    Another representative configuration of the mobile object detection apparatus according to the present invention comprises an input unit for introducing a mobile image, a frame acquisition unit which continuously extracts a plurality of images from raster from a moving image, a horizon line detection unit which detects a horizon line in a raster image, a rim image generation unit which generates an rim image from a frame image, and a moving object detection unit which defines a detection box on a moving object, and the edge image generation unit extracts a edge image which is connected to the horizon line detected by the horizon line detection unit, and the moving object detection unit converts a raster image to a grayscale image, extracts a difference between the grayscale image and an image background of the image in levels of g laugh using a background difference process and make the difference a foreground, and filter the foreground using the edge image which is connected to the horizon line.
    A configuration representative of a method of detecting a moving object according to the present invention comprises a step of introducing a moving image, a step of continuously extracting a plurality of frame images from of a moving image, a step of extracting an image of edges under a horizon line from a raster image, a step of extracting a difference between the image of edges under the line horizon and a background image of the edge image using a background difference method, a step of converting a raster image into a grayscale image and extracting a difference between the grayscale image and a background image of the grayscale image using the background difference method, and a step of generating a foreground by combining the difference between the edge image under the horizon line and the background image of the ledge image with the d difference (BDG) between the grayscale image and the background image of the grayscale image.
    According to the above method, it is possible to accurately detect a part in which the position in the image of a moving object is connected to the ground. So even if the object has a flat texture or is moving slowly, the whole object can be detected as the foreground, and the detection rate can be improved. Technical ideas similar to those of the moving object detection apparatus described above can be applied to the moving object detection method.
    Another configuration representative of a moving object detection method according to the present invention comprises a step of introducing a moving image, a step of continuously extracting a plurality of frame images from from a moving image, a step of extracting a rim image which is connected to a horizon line from a frame image, a step of converting a frame image into an image in grayscale, extracting a difference between the grayscale image and a background image of the grayscale image using a background difference method to make the difference a foreground, and a foreground filtering step using the edge image which is connected to the horizon line.
    According to the above method, the boundary between a building and the road can be removed as noise. In particular, when the moving object detection apparatus moves, the noise can be reduced significantly. Technical ideas similar to those of the moving object detection apparatus described above can be applied to the moving object detection method.
    Advantages provided According to the present invention, it is possible to accurately detect the position in the image of a part in which a mobile object is connected to the ground. Thus, even if the object has a flat texture, or is moving slowly, the entire object can be detected as the foreground, and the detection rate can be improved. Therefore, it is possible to provide a moving object detection apparatus and a moving object detection method which are inexpensive and safe, and to enable precise and high speed detection of a moving object from 'a moving image of a monocular camera.
    BRIEF DESCRIPTION OF THE DRAWINGS Other characteristics, details and advantages of the invention will appear on reading the detailed description below, and on analysis of the accompanying drawings, in which: Fig.l [0022] [ fig.l] is a diagram illustrating the general configuration of a mobile object detection apparatus according to this embodiment;
    Fig.2 [fig-2] is a flowchart illustrating a method of detecting a moving object according to this embodiment;
    Fig. 3 [fig.3] is a flowchart illustrating a procedure for detecting a moving object;
    Fig. 4 [fig-4] is a flowchart illustrating a procedure for extracting a foreground;
    Fig. 5 [fig.5] is an example image of a background difference process;
    Fig. 6 [fig.6] is an example of an image illustrating the generation of an image of edges (E);
    Fig. 7 [fig-7] is an example of an image illustrating the extraction of an image of ledges (BE) below the horizon line;
    Fig. 8 [fig.8] is an example of an image illustrating the extraction of a difference (DBE) between the image of ledges (BE) below the horizon line and the background;
    Fig. 9 [fig.9] is an example of an image illustrating the denoising of an image of edges;
    Fig. [Fig. 10] is an example of an image illustrating the combination of a difference (RBDBE) between the image of edges (BE) and a background image of the image of edges (BE);
    Fig. 11 [fig.l 1] is an example of an image illustrating an image of ledges (CE) which is connected to the horizon line;
    Fig. 12 [fig. 12] is an example of an image illustrating the filtering of a binarized image (BDG ’);
    Fig. 13 [fig. 13] is an example of an image illustrating enlargement and reduction of foregrounds;
    Fig. 14 [fig. 14] is an example of an image illustrating the definition of detection boxes;
    Fig. 15 [fig. 15] is another example of an image illustrating integration and separation;
    Fig. 16 [fig. 16] is an example of an image illustrating the processing by filtering of the detection boxes;
    Fig. 17 [fig. 17] is an example of an image illustrating the generation of a background image;
    Fig. 18 [fig. 18] is an example of an image illustrating the filtering treatment in which the horizon line is used;
    Fig. 19 [fig. 19] is a diagram illustrating the conditions and analogs of a genetic algorithm;
    Fig. [0041] [fig.20] is a flow chart of the genetic algorithm;
    Fig. 21 [fig.21] is a diagram illustrating an objective function of the genetic algorithm.
    Description of the embodiments A preferred embodiment of the present invention is described below in detail with reference to the accompanying drawings. The dimensions, materials, and other specific values indicated in the embodiment are given only for the purpose of illustration to facilitate understanding of the present invention and are not intended to limit the scope of the invention, except otherwise indicated. It should be noted that in the description and the drawings, the elements having substantially identical functions or structures are designated by identical reference signs to avoid any redundant description thereof. In addition, items which are not directly related to the present invention are omitted.
    An embodiment of a mobile object detection apparatus and a method of detecting a mobile object according to the present invention are described. Fig. 1 is a diagram illustrating the general configuration of a mobile object detection apparatus 100 according to this embodiment. Fig. 2 is a flowchart illustrating a method of detecting a moving object according to this embodiment.
    As shown in FIG. 1 (a), the mobile object detection device 100 is an on-board system installed in an automobile 10. As a representative example, a moving image taken by an on-board camera 12 of the automobile 10 is introduced into the the moving object detection apparatus 100, and a moving object is detected. Information about the moving object detected by the moving object detection apparatus 100 is sent to and used by a braking system 14, a screen (not shown), and the like.
    The on-board camera 12 is an inexpensive monocular camera, and takes an ordinary moving image. The moving image can be a color image or a grayscale image (monochrome image). In addition, the on-board camera 12 can also be fitted with a hypergone lens or a wide-angle lens. According to the present invention, as will be described later, a moving object can be accurately detected at high speed, even from a moving image of a monocular camera, and it is thus possible to provide a detection apparatus low cost and secure moving object detection method and moving object detection method.
    FIG. 1 (b) shows the configuration of the mobile object detection apparatus 100. The mobile object detection apparatus 100 can be constructed specifically with a computer system. The configuration of the mobile object detection apparatus 100 is described below with reference to the flowchart shown in Fig. 2.
    A moving image taken by the on-board camera 12 is introduced into an input unit 110 (step S210). If the moving image is a video image signal, a video encoder chip provided with a composite terminal or an HDMI terminal (registered trademark) corresponds to the input unit 110. If the moving image is encoded digital data, a USB port or an IEEE1394 interface corresponds to the input unit 110. In both cases, it is sufficient that the input unit 110 can introduce data of a moving image so that the data can be processed by the mobile object detection apparatus 100.
    A frame acquisition unit 112 continuously receives a plurality of frame images (still images) from a moving image which has been introduced (step S212). The specific processing for the acquisition of frames from a moving image depends on the format of the moving image. For example, if the moving image is in a format in which still images are arranged like in MotionJPEG, simply extract the frames. Note that the frames can be extracted at fixed intervals (for example, 0.1s) independently of the fps (frames per second, from the English "frames per second") of the moving image. For a moving image subjected to differential compression such as MPEG, I frames of a GOP (group of pictures, from English "group of pictures") can be extracted at the same time. A preprocessing unit 114 performs preprocessing on the frame images obtained (step S214). In this embodiment, the moving image that is introduced is a color image, and is converted to a grayscale image as a preprocessing. It is noted that, as a pretreatment, denoising can also be carried out, or a cropping treatment to eliminate a region to be subjected to image processing can also be carried out, if necessary.
    When a frame image to be processed is defined as a current frame, a moving object detection unit 118 detects a moving object from the current frame using the background difference method, and defines a detection box for the object (step S216). Note that, in the description which follows, a frame preceding the frame to be processed is called “previous frame”, and the subsequent frame is designated by “next frame”.
    Fig. 3 is a flowchart illustrating a procedure (subroutine) for detecting a moving object and FIG. 4 is a flowchart illustrating a procedure for extracting a foreground (step 232).
    As shown in FIG. 4, the mobile object detection unit 118 extracts a difference (DG) between a grayscale image, which is the current frame, and a background image of the grayscale image (step S250). Fig. 5 is an example of a background difference process image. In the image of the current frame shown in FIG. 5 (a), a horizon line is in the center of the image, and a pedestrian and a building are present on the right side. In addition, there are bushes and a fence along the horizon line. In this image, the pedestrian is a moving object.
    E background image shown in FIG. 5 (b) is obtained from a background image generation unit 116 described later. In addition, in the background image, a moving object is shown, but the position thereof gradually changes. If the background difference method is applied to these images (a luminance subtraction is carried out), pixels which indicate the difference are extracted, as in the difference image represented in FIG. 5 (c). The pixels which indicate the difference constitute the foreground.
    Ee foreground extracted using the background difference method has a low luminance and is in a finely fragmented state. Given this, smoothing (blurring processing) using a Gaussian filter or the like is performed so as to couple close foregrounds, after which a binarized image (BDG) is generated for clarification (step S252).
    The mobile object detection unit 118 performs processing of an edge image (E) in parallel with the processing of the grayscale image. Fig. 6 is an example of an image illustrating the generation of an image of edges (E). A rim image generation unit 122 generates the rim image (E) shown in FIG. 6 (b) (step S254) by setting to 1 if the pixel value of the difference with one of the adjacent pixels in the frame image shown in FIG. 6 (a) is greater than or equal to a threshold value, and by setting to 0 if the pixel value is less than the threshold value. The generation of a border image is also called "contour extraction processing".
    Next, the edge image generation unit 122 extracts an edge image (BE) below the horizon line (step S256). The horizon line is detected from the image of the current frame by a horizon line detection unit 124 using a genetic algorithm. The processing for detecting the horizon line is described below.
    [0058] FIG. 7 is an example of an image illustrating the extraction of the image of ledges (BE) below the horizon line. A horizon line HL detected by the horizon line detection unit 124 is drawn in the edge image (E) shown in FIG. 7 (a). The image of ledges (BE) below the horizon line HL is then extracted (step S258), as illustrated in FIG. 7 (b).
    [0059] FIG. 8 is an example of an image illustrating the extraction of the difference (DBE) between the image of ledges (BE) below the horizon line and the background image of the image of ledges (BE). The background image generation unit 116 generates the background image shown in FIG. 8 (b), using the image of edges (BE) below the horizon line shown in Fig. 8 (a). The edge image generation unit 122 then obtains the difference (DBE) between the edge image (BE) and the background image of the edge image (BE), as illustrated in FIG. 8 (c).
    The edge image is raw, and therefore the difference (DBE) between the edge image (BE) and the background image of the edge image (BE) also contains a lot of noise. Fig. 9 is an example of an image illustrating the denoising of an image of edges. The edge image generation unit 122 performs binarization processing on the difference (DBE) between the edge image (BE) and the background image of the edge image (BE) shown in FIG. . 9 (a), and generates a binarized image (BDBE), as illustrated in FIG. 9 (b). An enlargement and a reduction are also carried out, and a denoised difference (RBDBE) between the image of ledges (BE) below the horizon line (HL) and the background image of the image of ledges (BE) , as shown in Fig. 9 (c), is obtained (step S260). The enlargement and reduction are described later with reference to Fig. 13.
    The moving object detection unit 118 combines, with the binarized image (BDG), which is a grayscale image, the denoised difference (RBDBE) between the edge image (BE) under the horizon line (HL) and the background image of the edge image (BE). It should be noted that, although denoising has been carried out, the possibility exists that the difference (RBDBE) between the edge image (BE) and the background image of the edge image (BE) still contains a lot of noise.
    In view of this, the mobile object detection unit 118 determines the noise level of the denoised difference (RBDBE) between the image of ledges (BE) below the horizon line (HL) and l background image of the edge image (BE) (step S262). Fig. 10 is an example of an image illustrating the combination of a difference (RBDBE) between the edge image (BE) and a background image of the edge image (BE). As shown in Fig. 10, if the difference (RBDBE) between the edge image (BE) and the background image of the edge image (BE) contains a lot of noise (YES in step S262), the difference (RBDBE) is not combined, and the binarized image (BDG), which is a grayscale image, is used as such as a binarized image (BDG ') (step S264). If the difference (RBDBE) between the edge image (BE) and the background image of the edge image (BE) contains only little noise (NO in step S262), an image obtained by combining , with the binarized image (BDG), which is a grayscale image, the difference (RBDBE) between the edge image (BE) and the background image of the edge image (BE) is used as a binarized image (BDG ') (step S266).
    According to the above configuration, it is possible to accurately detect the position in the image of a part in which a mobile object is connected to the ground. Thus, if the object has a flat texture, or is moving slowly, the entire object can be detected as the foreground, and the detection rate can be improved. Therefore, it is possible to accurately detect a moving object from a moving image of a monocular camera at high speed, and to improve security at low cost.
    The binarized image (BDG ’) generated as described above can also be defined as foreground (EF) in this state. However, in this embodiment, in order to further improve the detection rate, the foreground is filtered using the image of edges which is connected to the horizon line.
    In the above description, when generating the edge image (E) (step S254), the mobile object detection unit 118 extracts the edge image (BE) below the line horizon (step S256). At the same time, an image of ledges (CE) which is connected to the horizon line is extracted from the image of ledges (E) (step S270).
    [0066] Fig. 11 is an example of an image illustrating the image of ledges (CE) which is connected to the horizon line. Fig. 11 (a) shows the same image of edges (E) as in Figs. 6 (b) and 7 (a). In this edge image (E), if only the pixels which are connected to the horizon line detected by the horizon line detection unit 124 are extracted, it is possible to obtain an edge image (CE ) which is connected to the horizon line as illustrated in Fig. 11 (b). Fig. 11 (c) shows an example of a conditional expression to extract the pixels.
    The moving object detection unit 118 then determines whether the vehicle 10 is moving or not, in other words, if the moving object detection apparatus 100 is moving or not (step S272) . If the moving object detection apparatus moves, the image positions of the road and the building change, and therefore the positions of their contours deviate, and ledges are detected by the background difference method. The border between the road and the building is not a moving object, and therefore these edges are noise.
    Eig. 12 is an example of an image illustrating filtering of the binarized image (BDG ’). As shown in Eig. 12, if the automobile 10 does not move (NO at step S272), the binarized image (BDG ’) is defined as being the foreground (EF) as it is (step S276). If the vehicle 10 moves (YES in step S272), filtering is carried out by performing an AND operation on the binarized image (BDG ') and the edge image (CE) which is connected to the line of horizon (only common pixels are kept) (step S274).
    According to the above configuration, the edge of the border between the road and the building which appears when the moving object detection device is moving can be eliminated as noise. It is therefore possible to reduce the noise significantly, and, in addition, it is possible to further improve the detection rate of a moving object.
    Next, the mobile object detection unit 118 performs denoising on the foreground (EF) (step S278). Fig. 13 is an example of an image illustrating enlargement and reduction of foregrounds. There are still voids and fine fragmentation, and therefore the foregrounds are enlarged as shown in Fig. 13 (b), and, subsequently, the foregrounds are reduced as illustrated in FIG. 13 (c) (step S236).
    Magnification is a process in which, when background pixels are defined as being black, the foreground pixels are defined as being white, and a filter diameter is defined as, for example, 5 × 5, the pixels inside the filter diameter centered on all the white pixels are changed to white. The reduction is opposite to the enlargement, and is a treatment to change the pixels inside the diameter of the filter centered on all the black pixels, in black. When enlargement and reduction are carried out, the outer edge contours return to their original position, but the voids and fragmentation (discontinuities) filled by enlargement remain to be filled, even if a reduction is carried out, and thus smoothing with close pixels can be performed.
    Here, as described at the start, in the background difference method, if an object has a flat texture, or the color of part of an object is similar to the background color, there is no no difference in luminance between the background image and the image of the current frame, and therefore there are cases where only part of the object can be detected. In this case, the foregrounds are separated from each other, and thus, the foregrounds cannot be coupled to each other only through the smoothing, enlargement and reduction described above. In view of this, in the present invention, the candidate boxes are integrated and separated, as described below.
    Integration is a treatment of what should in reality be just a moving object determined to have been extracted into a plurality of detection boxes due to a false detection, and the detection boxes are converted into a single detection box. Separation is a processing for the determination of nearby detection boxes as a plurality of nearby moving objects, and for their use as independent detection boxes.
    Lig. 14 is an example of an image illustrating the definition of detection boxes. The mobile object detection unit 118 detects the contours of the extracted foregrounds, and defines frames which include the foregrounds as primary candidate boxes 150a to 150f as illustrated in Fig. 14 (a) (step S238). The mobile object detection unit 118 then obtains the detection boxes of the previous frame from a storage unit 120 (step S240). Note that if there is still no detection box of the previous frame in the current frame, this integration and separation processing (steps S240 to S244) is ignored.
    The mobile object detection unit 118 then defines, among the primary candidate boxes 150a to 150f shown in Lig. 14 (a), primary candidate boxes which overlap detection boxes 152a to 152c of the previous frame on a predetermined area or more (e.g. 50% or more), such as secondary candidate boxes 154a to 154d, as illustrated in Lig. 14 (b) (step S242). In Lig. 14 (b), the detection boxes 152 of the previous frame are indicated by dashed lines, and the secondary candidate boxes 154 are indicated by dotted lines. Primary candidate boxes 150e and 150f which have not become secondary candidate boxes, among primary candidate boxes 150a to 150f, are considered to be detection boxes as such (a "detection box" is not a candidate).
    The mobile object detection unit 118 groups the secondary candidate boxes which overlap the detection boxes 152 of the previous frame, among the secondary candidate boxes 154. The secondary candidate boxes 154a and 154b overlap the detection box 152a from the previous frame positioned on the left of the Lig. 14 (b), and are therefore grouped together. Thereafter, the grouped secondary candidate boxes 154a and 154b, which overlap in the y-axis direction on a predetermined area or more (e.g. 30% or more), are defined as an integrated detection box 156a as illustrated in Lig. 14 (c). If the grouped candidate boxes 154a and 154b do not overlap in the direction of the y axis on the predetermined area or more, they are defined as separate (independent) detection boxes (step S244). Note that "overlap in the direction of the y axis on a predetermined surface" is synonymous with the ranges of the coordinates of the boxes in the direction of the x axis by overlapping on a predetermined surface.
    To organize the above, the following three types of boxes are considered to be "detection boxes".
    - primary candidate boxes which have not become secondary candidate boxes (150e and 150f)
    - secondary candidate boxes which have been integrated (156a)
    - secondary candidate boxes which remain separate instead of being integrated (154c and 154d) It is noted that if there is a large number of secondary candidate boxes grouped, cases exist in which a plurality of integrated detection boxes are defined. For example, if a group includes four secondary candidate boxes, there are cases in which two detection boxes each comprising two integrated secondary candidate boxes are defined.
    Fig. 15 is another example image illustrating integration and separation, and are examples of treatment when two pedestrians who appeared to overlap separate from each other. Fig. 15 (a) illustrates the image of the current frame. The two pedestrians walk on a road. Fig. 15 (b) illustrates first planes extracted from the current frame, and corresponds to a central part of the current frame. The foreground of the person on the right is represented as a primary candidate box 150g, while the foreground of the person on the left is represented as two primary candidate boxes 150h and 150i which are vertically separated. The number of foreground blocks of the two pedestrians is three in total. Fig. 15 (c) shows a foreground extracted from the previous frame and a detection box. The two pedestrians always seemed to overlap in the previous frame, and therefore the two people are represented as a large detection box 152d.
    Then, as shown in FIG. 15 (d), the moving object detection unit 118 superimposes the detection box 152d of the previous frame on the primary candidate boxes 150, and defines primary candidate boxes 150 which overlap the detection box 152d on a predetermined surface or more, as secondary candidate boxes. In this example, the three primary candidate boxes 150g to 150i all become three secondary candidate boxes 154e to 154g. In addition, the three secondary candidate boxes 154e to 154g overlap the detection box 152d of the same previous frame, and therefore these are grouped.
    In Fig. 15 (d), the secondary candidate boxes 154F and 154g overlap each other in the direction of the y axis on a predetermined surface or more, and are therefore defined as an integrated detection box 156b as illustrated in FIG . 15 (e). The secondary candidate box 154e does not overlap the integrated detection box 156b in the y-axis direction on the predetermined surface or more, and thus remains separate instead of being integrated. Consequently, as shown in Fig. 15 (f), the respective detection boxes (the secondary candidate box 154e and the integrated detection box 156b) are defined for the two pedestrians. Note that, when determining the degree of overlap, the size (length) in the x-axis direction of a part in which two candidate boxes overlap can be compared to the size in the x-axis direction of one of the two candidate boxes which is weaker in this direction.
    In this way, it is possible to appropriately integrate the first three planes, and to define detection boxes on the two moving objects. In addition, at the same time, a detection box in the previous frame can be separated into two detection boxes in the current frame.
    [0083] Fig. 16 is an example of an image illustrating the processing by filtering of the detection boxes (step S246). Fig. 16 (a) shows the detection boxes after completion of the integration and separation (see Fig. 14 (c). In Fig. 16 (a), the driver of a bicycle on the left of the image is defined as the built-in detection box 156a, and the two cars in the center of the image are detected as secondary candidate boxes 154c and 154d. In addition, although the primary candidate boxes 150e and 150f remain as detection, they are extremely small, and can therefore be considered as noise.
    In view of this, in this embodiment, a filter unit 126 eliminates the detection boxes which do not meet a predetermined surface or aspect ratio, from one or more detection boxes detected in the current frame (ASF: in English Aera Size Filtering or Filtering by surface size). Therefore, the extremely small detection boxes and the extremely long and thin detection boxes are eliminated.
    In the example of FIG. 16 (b), extremely small detection boxes (primary candidate boxes 150e and 150f) are suppressed by filtering processing, and the integrated detection box 156a and secondary candidate boxes 154c and 154d remain. Note that, in FIG. 16 (b), these final detection boxes are superimposed on the grayscale image of the current frame, and not a binary image, and are displayed.
    A moving image obtained by taking the shot while the shooting position is in motion, an image of a monocular camera, and, in addition, the use of a hypergone lens or of a wide-angle lens causes increased noise. In view of this, it is possible to eliminate the noise by carrying out the filtering processing described above, and to improve the precision of the detection of a moving object.
    The mobile object detection unit 118 records the final detection boxes on which the filter unit 126 has carried out the processing by filtering, in the storage unit 120 (step S218). Detection boxes are stored in association with frame numbers, and it is possible to read recorded detection boxes for an image of any frame. In particular, as the detection boxes of the previous frame were used for the image processing of the current frame in the description above, the detection boxes of the current frame will be used for the image processing of the next frame. In addition, the detection boxes which have been defined by the mobile object detection device 100 are transmitted to the braking system 14, a monitor (not shown), etc. by an output unit 130.
    The background image generation unit 116 generates (updates) a background image to process the image of the next frame (step S220). No foreground is extracted from the first frame image among a plurality of frame images obtained by the frame acquisition unit 112, and the first frame image is only used as a background image. When foregrounds are extracted from the second frame image, the first frame image is used as the background image. When foregrounds continue to be extracted from the third frame image, a background image generated (updated) by the background image generating unit 116 is used.
    [0089] Fig. 17 is an example of an image illustrating the generation of a background image. The background image generation unit 116 combines a background image and the image of the current frame to update the background image, and defines the updated background image as a background image. background when processing the next frame. As a specific example, as shown in Fig. 17, an image in which the luminance of the current frame is reduced to 25% and an image in which the luminance of a background image is reduced to 75% are combined, and a background image is produced. By sequentially generating (updating) a background image in this way, a background image can be generated, as the case may be, even in the case of a captured moving image while moving the camera, and a moving image which is never in a state in which there is no moving object.
    In addition, in the background difference method, the computation cost for an update of a background model which is based on learning is high, and therefore it has been considered difficult to apply. such an update from a background model to high speed mobile image processing. However, the amount of computation for combining a past background image and the image of the current frame as in this embodiment is small, and it is therefore possible to increase the efficiency relative to the computation cost of the entire system.
    When the generation of a background image is finished, it is determined whether there is a next frame image or not (step S222). If there is a next frame image, the series of processing described above is repeated. If there is no next frame, the procedure ends.
    As described above, according to the moving object detection apparatus and the moving object detection method according to this embodiment, the integration and the separation are carried out on candidate boxes in the current frame. , in the range of candidate boxes which overlap detection boxes of the previous frame. It is not necessary to determine the type or distance of a moving object, and it is also possible to integrate and separate candidate boxes even if there is a lot of noise.
    Consequently, even in the case of an image of an inexpensive monocular camera and of an image taken by a camera such as an on-board camera while it is moving, it is possible to define boxes sensing device corresponding appropriately to moving objects. Furthermore, the background difference method allows processing at high speed, and also the integration and separation processing of candidate boxes and the filtering processing which have been described above can be carried out at high speed. . Therefore, approaching moving objects can be accurately detected at a high speed, and security can be improved.
    In particular, the integration and separation algorithm is simple, and thus, compared to a simple conventional background difference method, the increase in the amount of calculations is small despite the large increase. the detection rate of a moving object. This technology is very practical, and its technical advantage is sufficiently high compared to conventional technologies. In addition, in the present invention, only a monocular camera is used, and the present invention is thus very advantageous in terms of costs compared to conventional technologies in which a plurality of cameras, a laser radar, an infrared camera, etc. are used.
    The present invention can be implemented in an on-board safe driving system to prevent frontal traffic accidents and in an automatic movement system of a mobile device such as a robot. In addition, no limitation exists as to the systems installed in automobiles, robots, etc., and the present invention can also be implemented in surveillance systems which are fixedly installed and use a wide angle security camera. .
    In the above embodiment, a description has been made according to which the filter unit 126 eliminates the detection boxes which do not meet a predetermined surface or aspect ratio (ASF). In this regard, a configuration can also be adopted in which, instead of or in addition to the ASF, the horizon line detection unit 124 detects the horizon line, and one or more detection boxes which do not overlap the horizon line according to a predetermined ratio, among one or more detection boxes detected in the current frame, are deleted (HFF: in English "Horizontal Fine Filtering", or Figne d'Horizontale filtering).
    Fig. 18 is an example of an image illustrating a filtering treatment using a horizon line. Fig. 18 (a) shows a binarized image and detection boxes (on which processing until integration and separation has been carried out). Fig. 18 (b) represents a detection box subjected to filtering processing and a grayscale image of the current frame on which the detection box is superimposed.
    As appears from the comparison of FIG. 18 (a) with FIG. 18 (b), in a moving image taken by an on-board camera or the like, while the camera is mobile, the possibility exists that a road sign and a pedestrian crossing (a detection box 160b of the pedestrian crossing) in a lower part of the image, and buildings, advertising signs, a traffic light, electric cables, public lighting, etc. (a 160c detection box for the advertising panel, a 160d detection box for the electric cable and a 160th detection box for the public lighting) in the upper part of the image are detected as moving objects. However, a mobile body which it is desired to detect according to the present invention is an object which moves on the ground such as a pedestrian or a vehicle (a detection box 160a of the vehicle).
    Taking this into account, as an expression of determination illustrated in FIG. 18 (c), zones A and B separated by the HF horizon line are obtained for all the detection boxes 160 and the detection boxes which do not overlap the HF horizon line according to a predetermined ratio are deleted. In this embodiment, the detection boxes for which the determination expression: la-bl / (a + b) <0.6 is true are kept (Correct = 1), and the other detection boxes are deleted (Correct = 0).
    As a specific example of the determination expression, since a = b when the horizon line HE passes through the center of a detection box, the value of the determination expression is equal to 0, Correct = 1, and the detection box 160 remains. When the horizon line HE does not overlap a detection box, the value of the determination expression is equal to 1, Correct = 0, and the detection box 160 is deleted.
    Note that the threshold value is set to 0.6, for example, and the numerical value of the threshold value can be determined if necessary. For example, the threshold value can also be set to la-bl / (a + b) <l. In this case, only a detection box that the HL horizon line does not overlap at all is deleted.
    In this way, it is possible to delete a detection box 160 which is not a detection box for a mobile unit, via appropriate processing and at high speed. Consequently, it is possible to increase the precision of the detection of a mobile object that one wishes to detect, in order to avoid a danger.
    Various technologies for the detection of a horizon line can be envisaged, but, in this embodiment, a genetic algorithm is used to detect a horizon line. Since the genetic algorithm is a multipoint search algorithm to improve a solution while maintaining a plurality of solutions which are always favorable locally, it is possible to search for the best solution globally, and therefore the genetic algorithm is the preferred method . Therefore, it is possible to accurately detect a horizon line that is difficult to easily recognize, hampered by the shapes of a building and the road.
    [0104] FIG. 19 is a diagram illustrating conditions and the like of a genetic algorithm. To briefly describe the detection of a horizon line to be described below, the boundary line between a road surface and another object (a structure such as a building, sky, or the like) is detected using the algorithm genetic, and is defined as a skyline. P ; (x i5 y ; ) designates an i-th pixel (point) on the horizon line. x ; is an x coordinate of the i-th point on the horizon line. y ; is a y coordinate of the i-th point on the horizon line, and is represented as the height of a baseline B + constant a + variable bi. In addition, a, b 0 , b b ... b M are defined as chromosomes of the genetic algorithm.
    [0105] Fig. 20 is a flow chart of the genetic algorithm. The filter unit 126 first initializes the algorithm using predetermined initial values (step S250). The initial values of the algorithm can be defined as the number of generations (50), the number of initial individuals (50), the crossover rate (0.7), and the mutation rate (0.01) , for example.
    Then, temporary parameters are assigned to initial individuals, and a candidate line as the horizon line is generated (step S252). While repeating the cross-checking and the mutation on the basis of the logic of the genetic algorithm, the degree of correspondence to an objective function of the candidate line is evaluated (step S254). Note that the overlap and mutation in the genetic algorithm are considered to be a known technology, and therefore a detailed description of them is omitted here.
    [0107] FIG. 21 is a diagram illustrating an objective function of the genetic algorithm. An objective function F represents the number of pixels Diff (j, x) whose difference in pixel values exceeds a threshold value, j denotes an evaluation width, and an evaluation range L is the maximum value of j. x denotes an x coordinate of an intermediate position between the pixel P, (x i5 y,) and a pixel Ρ, + Ι (x, + l, y ; +1). Y x denotes a function of x which represents the y coordinate of a pixel P by linear interpolation (simple equation).
    [0108] The pixel value (grayscale value) of the i-th pixel P (χ;, γ;) is expressed in the form P (χ;, γ;). Regarding the determination of Diff pixels (j, x) whose pixel values are different, 1 is defined if the difference in pixel value IP (x, y x + j) -P (x, y x -j ) l is greater than or equal to a threshold value, and 0 is defined if it is less than the threshold value. DiffTB (j, x) equals 1 if the pixel values (grayscale value) are different from a threshold value or more as a result of comparing pixel values at separate positions in the y-axis direction by 2d. Consequently, the value of the objective function F is greater when the horizon line formed by the chromosomes a, b 0 , b b ... b M is closer to the delimitation line between the road surface and another object. In other words, the detection of the horizon line using this genetic algorithm is an extraction of a contour from the upper edge of the road surface.
    In addition, DiffTN and DiffBN in the evaluation function F are respectively decided as the results of the comparison of the pixel values of the pixels P (x, y x -j) and P (x, y x + j) and of those adjacent to it (next to these). Consequently, the value of the evaluation function f also increases when not only a horizontal line but also an outline (vertical line) of a building or the like is visible. The value is higher at the intersection between the horizontal line and the vertical line, in other words, the value is higher at the intersection between the building and the ground. Therefore, it is possible to improve the accuracy of the horizon line detection.
    When the calculation of a predetermined number of generations ends, the filter unit 126 outputs a candidate line selected as being the horizon line (step S256). The horizon line which has been delivered is used for the expression of determination described with reference to FIG. 18 (c).
    The filter unit 126 then determines whether there is a next frame or not (step S258), and if there is a next frame, repeats the detection of the horizon line using the algorithm genetic, and if there is no next frame, ends the processing for detecting the horizon line.
    The detection of a horizon line using the genetic algorithm can be performed on an image. In this case, the processing can be carried out on all the frames, but it is possible to increase the speed significantly by carrying out the processing on frames at each fixed time. In addition, by performing the processing at any fixed time when there is no change in the image, and by performing the processing also when the image largely changes, it is possible to increase the speed processing and improve accuracy.
    Furthermore, if an acceleration sensor is installed, or if a signal from an acceleration sensor is received from a navigation system, the processing can also be carried out when the camera takes a turn or the direction of movement changes between up and down. Therefore, it is easy to follow the change of the horizon line, and the detection accuracy can be improved.
    In addition, with regard to the detection of a horizon line using the genetic algorithm, it is possible to increase the speed and improve the precision by using a result of calculation of the previous frame. (the image of the previous frame is not used). More precisely, a calculation is performed using chromosome information from all the individuals of the genetic algorithm which includes the optimal solution calculated in the previous frame, as precursor chromosomes of a population of initial individuals (processing of evolving moving image). In particular, during the execution of the processing on all the frames, or during the execution of the processing on the frames defined at short time intervals, the displacement of the horizon line in the image is small, and thus, if the result of the calculation of the previous frame is used, a conversion to the optimal solution takes an extremely short time. Consequently, even a CPU (Central Processing Unit) of an on-board computer can perform the calculations at a speed sufficient for processing in real time, and it is possible to detect continuously the horizon line.
    As described above, by eliminating the detection boxes which do not overlap the horizon line according to a predetermined ratio, it is possible to increase the precision of the detection of a moving object. In addition, by using the genetic algorithm for the detection of a horizon line, it is possible to accurately detect the horizon line.
    Industrial Application [0116] The present invention can be used as a moving object detection apparatus and moving object detection method which are inexpensive and safe, and can accurately detect a moving object from a moving image d '' a high speed monocular camera.
    List of reference signs [0117] - 10 ... vehicle;
    - 12 ... on-board camera;
    - 14 ... braking system;
    - 100 ... mobile object detection device;
    - 110 ... input unit;
    - 112 ... frame acquisition unit;
    - 114 ... pretreatment unit;
    - 116 ... background image generation unit;
    - 118 ... mobile object detection unit;
    - 120 ... storage unit;
    - 122 ... edge image generation unit;
    - 124 ... horizon line detection unit;
    - 126 ... filter unit;
    - 130 ... output unit;
    - 150 ... primary candidate box;
    - 152 ... previous frame detection box;
    - 154 ... secondary candidate box;
    - 156 ... integrated detection box;
    - 160 ... detection box
    List of documents cited
    Patent documents [0118] For all practical purposes, the following patent document is cited:
    - [Patent Document 1] Publication of Japanese unexamined patent application 2010-092353.
    权利要求:
    Claims (1)
    [1" id="c-fr-0001]
    [Claim 1] [Claim 2] [Claim 3]
    Claims
    A moving object detection apparatus (100) comprising: an input unit (110) for inputting a moving image; a frame acquisition unit (112) which continuously extracts a plurality of frame images from a moving image;
    a horizon line detection unit (124) which detects a horizon line (HL) in a frame image;
    a rim image generation unit (122) which generates a rim image (E) from a frame image; and a moving object detection unit (118) which defines a detection box (152a, 152b, 152c, 152d) on a moving object, wherein the edge image generating unit (122) extracts an image edges (BE) below the horizon line (HL) detected by the horizon line detection unit (124), and extracts a difference (RBDBE) between the image of edges (BE) below the line d horizon (HL) and a background image of the edge image (BE) using a background difference method, and the moving object detection unit (118) converts a frame image into an image grayscale and extracts a difference (BDG) between the grayscale image and a background image of the grayscale image using the background difference method, and generates a foreground (EF) by combining the difference (RBDBE) between the edge image (BE) below the horizon line (HL) and the background image of the edge image (BE) with the difference (BDG) between l Image grayscale and the background image of the image to grayscale. The moving object detection apparatus (100) according to claim 1, wherein the moving object detection unit (118) combines the difference (RBDBE) between the image of ledges below the horizon line ( HL) and the background image of the edge image (BE) with the difference (BDG) between the grayscale image and the background image of the grayscale image, if the difference ( RBDBE) between the edge image (BE) below the horizon line (HL) and the background image of the edge image (BE) contains little noise.
    The moving object detection apparatus (100) according to claim 1 or 2, wherein the ledge image generating unit (122) extracts a ledge image (CE) which is connected to the line of sight. horizon (HL) detected
    by the horizon line detection unit (124), andthe moving object detection unit (118) filters the foreground (EF) using the edge image (CE) which is connected to the horizon line (HL). [Claim 4] The moving object detection apparatus (100) according to claim 3, wherein, when the moving object detection apparatus (100) moves, the moving object detection unit (118) filters the foreground (EF) using the edge image (CE) which is connected to the horizon line (HL). [Claim 5] A moving object detection apparatus (100) comprising: an input unit (110) for inputting a moving image;a frame acquisition unit (112) which continuously extracts a plurality of frame images from a moving image;a horizon line detection unit (124) which detects a horizon line (HL) in a frame image;a rim image generation unit (122) which generates a rim image (E) from a frame image; anda mobile object detection unit (118) which defines a detection box (152a, 152b, 152c, 152d) on a mobile object,wherein the edge image generation unit (122) extracts a edge image (CE) which is connected to the horizon line (HL) detected by the horizon line detection unit (124) , andthe moving object detection unit (118) converts a raster image to a grayscale image, extracts a difference (BDG) between the grayscale image and a background image of the grayscale image of gray using a background difference process and making the difference (BDG) a foreground (EF), andfilters the foreground (EF) using the edge image (CE) which is connected to the horizon line (HL). [Claim 6] The moving object detection apparatus (100) according to claim 5, wherein, when the moving object detection apparatus (100) moves, the moving object detection unit (118) filters the foreground (EF) using the edge image (CE) which is connected to the horizon line (HL). [Claim 7] A method of detecting a moving object comprising: a step of introducing a moving image;a step of continuously extracting a plurality of frame images from a moving image;a step of extracting an image of edges (BE) under a line
    [Claim 8] of horizon (HL) from a frame image;
    a step of extracting a difference (RBDBE) between the edge image (BE) below the horizon line (HL) and a background image of the edge image (BE) using a difference method background ; a step of converting a raster image into a grayscale image and extracting a difference (BDG) between the grayscale image and a background image of the grayscale image in using the background difference method; and a step of generating a foreground (EF) by combining the difference (RBDBE) between the edge image (BE) below the horizon line (HL) and the background image of the edge image (BE) with the difference (BDG) between the grayscale image and the background image of the grayscale image.
    A method of detecting a mobile object comprising:
    a step of introducing a moving image;
    a step of continuously extracting a plurality of frame images from a moving image;
    a step of extracting an edge image (CE) which is connected to a horizon line (HL) from a frame image;
    a step of converting a raster image into a grayscale image, extracting a difference (BDG) between the grayscale image and a background image of the grayscale image in using a background difference process to make the difference (BDG) a foreground (EF); and a foreground filtering step (EF) using the edge image (CE) which is connected to the horizon line (HL).
    类似技术:
    公开号 | 公开日 | 专利标题
    Börcs et al.2017|Instant object detection in lidar point clouds
    US9378424B2|2016-06-28|Method and device for detecting road region as well as method and device for detecting road line
    JP5867807B2|2016-02-24|Vehicle identification device
    FR3085219A1|2020-02-28|MOVING OBJECT DETECTION APPARATUS AND MOVING OBJECT DETECTION METHOD
    Creusot et al.2015|Real-time small obstacle detection on highways using compressive RBM road reconstruction
    FR2914761A1|2008-10-10|DEVICE FOR RECOGNIZING AN OBJECT ON AN IMAGE
    FR2674354A1|1992-09-25|Method of analysing sequences of road images, device for implementing it and its application to obstacle detection
    Anand et al.2018|Crack-pot: Autonomous road crack and pothole detection
    JP2006318059A|2006-11-24|Apparatus, method, and program for image processing
    JP2006318060A|2006-11-24|Apparatus, method, and program for image processing
    EP2930659B1|2016-12-21|Method for detecting points of interest in a digital image
    Zhang et al.2020|ISSAFE: Improving semantic segmentation in accidents by fusing event-based data
    Goga et al.2018|Fusing semantic labeled camera images and 3D LiDAR data for the detection of urban curbs
    CN110889403A|2020-03-17|Text detection method and related device
    EP0588815A1|1994-03-30|Road image sequence analysis device and method for obstacle detection
    Pavlov et al.2019|Icevisionset: lossless video dataset collected on russian winter roads with traffic sign annotations
    JP2006318061A|2006-11-24|Apparatus, method, and program for image processing
    JP2006317193A|2006-11-24|Image processing apparatus, imaging processing method and program for processing image
    Barua et al.2019|An Efficient Method of Lane Detection and Tracking for Highway Safety
    FR3074941A1|2019-06-14|USE OF SILHOUETTES FOR RAPID RECOGNITION OF OBJECTS
    WO2018138064A1|2018-08-02|Detection of obstacles in the environment of a motor vehicle by image processing
    JPH11231069A|1999-08-27|Evaluation method for rain and snow falling condition and its device
    WO2017005930A1|2017-01-12|Detection of objects by image processing
    FR2858447A1|2005-02-04|AUTOMATED PERCEPTION METHOD AND DEVICE WITH DETERMINATION AND CHARACTERIZATION OF EDGES AND BORDERS OF OBJECTS OF A SPACE, CONSTRUCTION OF CONTOURS AND APPLICATIONS
    JP7018607B2|2022-02-14|Moving object detection device and moving object detection method
    同族专利:
    公开号 | 公开日
    DE102019122690A1|2020-02-27|
    US11024042B2|2021-06-01|
    US20200065981A1|2020-02-27|
    JP2020030768A|2020-02-27|
    引用文献:
    公开号 | 申请日 | 公开日 | 申请人 | 专利标题
    US6430303B1|1993-03-31|2002-08-06|Fujitsu Limited|Image processing apparatus|
    US7227893B1|2002-08-22|2007-06-05|Xlabs Holdings, Llc|Application-specific object-based segmentation and recognition system|
    US7764808B2|2003-03-24|2010-07-27|Siemens Corporation|System and method for vehicle detection and tracking|
    US7668376B2|2004-06-30|2010-02-23|National Instruments Corporation|Shape feature extraction and classification|
    TW201001338A|2008-06-16|2010-01-01|Huper Lab Co Ltd|Method of detecting moving objects|
    JP5083164B2|2008-10-09|2012-11-28|住友電気工業株式会社|Image processing apparatus and image processing method|
    TWI393074B|2009-12-10|2013-04-11|Ind Tech Res Inst|Apparatus and method for moving object detection|
    JP5999043B2|2013-07-26|2016-09-28|株式会社デンソー|Vehicle periphery monitoring device and program|
    US10091507B2|2014-03-10|2018-10-02|Euclid Discoveries, Llc|Perceptual optimization for model-based video encoding|
    JP6550881B2|2014-07-14|2019-07-31|株式会社リコー|Three-dimensional object detection device, three-dimensional object detection method, three-dimensional object detection program, and mobile device control system|
    US10089549B1|2015-08-27|2018-10-02|Hrl Laboratories, Llc|Valley search method for estimating ego-motion of a camera from videos|
    US10078902B1|2015-08-27|2018-09-18|Hrl Laboratories, Llc|Fast robust method for compensating ego-translations and detecting independent moving objects in video captured with a moving camera|
    WO2017077261A1|2015-11-05|2017-05-11|Quantum Vision Technologies Ltd|A monocular camera cognitive imaging system for a vehicle|
    JP6208277B2|2016-03-31|2017-10-04|大王製紙株式会社|Packaging bag|CN111553282A|2020-04-29|2020-08-18|北京百度网讯科技有限公司|Method and device for detecting vehicle|
    CN113450386B|2021-08-31|2021-12-03|北京美摄网络科技有限公司|Face tracking method and device|
    法律状态:
    2020-06-24| PLFP| Fee payment|Year of fee payment: 2 |
    2021-07-29| PLFP| Fee payment|Year of fee payment: 3 |
    优先权:
    申请号 | 申请日 | 专利标题
    JP2018-157668|2018-08-24|
    JP2018157668A|JP2020030768A|2018-08-24|2018-08-24|Moving object detecting device and moving object detecting method|
    [返回顶部]