西班牙专利ES2563098A1 Ir image improvement procedure based on scene information for video analysis (Machine-translation by

专利PDF首页>>西班牙专利

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:
Ir image improvement procedure based on scene information for video analysis. Image improvement procedure for video analysis or automatic video surveillance systems comprising at least one image acquisition device through which an image in the ir or thermal spectrum of an area of the space is captured, a scene and scene calibration system a detection system through which at least one type of object is detected, said method comprising at least one processing step in which the contrast of the image captured by the image acquisition device is improved through the information depth or scene of the image, obtained, directly or indirectly, by the scene calibration system or manually entered by the user. (Machine-translation by Google Translate, not legally binding)
公开号:ES2563098A1
申请号:ES201530836
申请日:2015-06-15
公开日:2016-03-10
发明作者:Nicolás HERRERO MOLINA；Martí BALCELLS CAPELLADES；Jordi LLUIS BARBA
申请人:Davantis Tech Sl；Davantis Technologies Sl；
IPC主号:

专利说明:

IR image enhancement procedure based on scene information for video analysis. 5 Field of technique
The present invention refers to an image improvement procedure for video analysis or automatic video surveillance systems whose input image is IR or thermal spectrum, in which depth or scene information is used for
10 make this improvement. Prior State of the Art
A generic video analysis system aims to determine the presence of
15 people, vehicles or other specific objects (objective) in a given area of space from the images captured by an image acquisition device, preferably, a fixed camera, which observes said particular area of space.
In the event that the input image to the aforementioned video analysis system is of
20 IR (infrared) or thermal spectrum, one of the main problems that this system must face is the lack of contrast between the scene (background) and the target (foreground), making it difficult to detect the target. For example, for the thermal case, at certain times of the year, mainly in the warmest ones, the soil temperature can be around 37 ° C, which can make it difficult to detect human targets by
25 the lack of contrast of the image. This effect is further accentuated if it fits in the remote areas of the image where objects are smaller. This problem not only suffer from automatic surveillance or video analysis systems but also those verified by an operator.
30 The manufacturers of infrared spectrum cameras have made an effort to add software tools to their sensors that reduce the contrast problem and, consequently, improve the image quality. The algorithms developed for this purpose are those that allow you to adjust the brightness or gain or perform a basic histogram equalization or equalizations
35 complex plateau type or detail enhancement filters. For their part, the manufacturers of video analysis systems, aware of this problem, have also

Built-in image enhancement and / or contrast enhancement modules to mitigate it.
The drawback of the aforementioned algorithms is that to make their improvements they rely
5 only and exclusively in the information present in the image. In other words, they treat all pixels equally without making any assumption about their nature (eg from what distance the projected temperature comes or what is the size of a person in that place).
10 However, it should be noted that in the scientific literature there are documents that describe procedures or methods of image enhancement that are based on scene and / or depth information in which mostly visible spectrum cameras are used. Among these documents the most significant are the following:
15 -On the one hand we find references, such as the patent application with publication number US2010 / 0304854 or in the document of Hachicha et al., In “20th European Signal Processing Conference” (EUSIPCO 2012) and with title “Combining Depth Information and Local Edge Detection for Stereo Image Enhancement ”, in which the objective is not to improve image quality
20 using a depth reference but improve said depth estimate based on the analyzed image. In short, the reverse process noted above is performed.
o The first of the documents cited as an example describes a system that calculates image depth based on emitting structured light and analyzing
25 its projection on the scene. In this document, an attempt is made to improve the quality of the projected structured light image to improve depth estimation.
o In the second of the documents cited as an example we have a stereo pair (visible image + depth map) and we try to segment the 30 objects of the image, improve their contrast and apply this information to improve the depth map.
- On the other hand we find publications, such as the document by Hu et al., In ICIP 2013 and entitled “KINECT DEPTH MAP BASED ENHANCEMENT FOR LOW LIGHT SURVEILLANCE IMAGE” and the patent application with the number of
35 publication CN103400351A, in which information from external devices (for example: Kinect camera) is applied to visible spectrum cameras. These

devices give us a depth map of the scene associated with the image of the scene that is being observed with the visible spectrum camera to perform enhancement operations on said visible spectrum camera. Specifically, in the first case a noise filtering and a histogram adjustment are performed in which
5 incorporates the depth image into the overall equalization of the image.
- There are also a large number of references, such as the paper by Srinivasa et al., In "IEEE TransactionsOnPatternAnalysis and Machine Intelligence, vol. 25, No. 6, June 2003 ”, entitled“ Contrast Restoration of Weather Degraded Images ”or the document of Munira et al., In“ International Journal of Scientific
10 and Research Publications, Volume 3, Issue 6, June 2013 ”, entitled“ Single Image Fog Removal Using Depth Estimation Basedon Blur Estimation ”, which aims to improve the quality of images that are affected by weather conditions (eg fog or pollution ), estimating the depth of the scene based precisely on these effects. Specifically, the more blurred
15 is an area of the image means that the camera is at a greater distance. Once the depth estimate is obtained, it is used to cancel the effects produced by the atmospheric conditions on the image, thus improving its quality. - Another group of documents must be highlighted, such as: the patent
20 with publication number US7,706,576 B1, in which objects of interest are detected to improve the quality of the objects themselves. In the document cited as an example, it is assumed that a person's face exists on the scene, that face is detected and an image improvement algorithm is applied in the region that occupies it. Actually, scene information is being applied to improve so
25 minus one area of the image.
As for scientific documents on procedures and methods for improving IR or thermal images, the most significant documents within the few existing references are the patent application with publication number EP 2226762A1 and the
30 PCT patent application with publication number WO 2013126001A2, use global or local information of the image itself and in no case depth or scene information is used. The main drawback of these procedures is that in the images there are still areas where there is not enough contrast to detect objects.

The present invention has as its main purpose to describe an image improvement procedure for video analysis or automatic video surveillance systems whose input image is IR or thermal spectrum in which, using depth or scene information, the resulting image has sufficient contrast to determine the presence of a
5 specific object. Explanation of the invention.
The present invention manages to overcome all the drawbacks as well as defects 10 mentioned above in the state of the art and achieve the purpose described in the previous paragraph.
Video analysis or automatic surveillance systems comprise at least one image acquisition device through which different images of a particular area of space are obtained which together with an image digitization system provides the output of the same digital image.
Then said image is subjected to an image processing system that applies at least one image improvement procedure such that at its output 20 said image must be of sufficient quality to detect a specific type of object. (person, animal or any other user defined).
Specifically, video analysis or automatic surveillance systems offer at least two modes of operation, the system calibration and detection. Such modes of operation are performed by their corresponding calibration and detection systems.
The calibration mode of operation is generally used at the beginning of the commissioning of the video analysis or automatic surveillance system since its purpose 30 is to provide the image with a spatial reference so that the detection system can reference all calculations What you do during the detection procedure: calculation of distance traveled, speed, size of objects, etc. Specifically, the calibration stage performed by the calibration system must allow the equivalence between the approximate size in pixels of the object to be detected (usually person) and each
35 one of the pixel coordinates of the image.

Well, as mentioned above, an image improvement procedure for video analysis or automatic video surveillance systems whose input image is IR or thermal spectrum in which depth information is used for this purpose is object of the present invention or image scene. To do this, the first step of this
5 procedure is that said depth or scene information of the image is entered by the user or that it comes from the calibration stage or the calibration system since obtaining the variation of the approximate size of the object to be detected for each of the pixel coordinates of the image is an indirect form of the depth of the actual scene captured in the image.
10 In this regard, it should be noted that for this purpose any calibration procedure that allows obtaining such results can be used, the most representative being the following: - Strong calibration procedure based on obtaining both the
15 intrinsic camera parameters (focal length, pixel size, radial lens distortion, etc.) and extrinsic parameters (height and angular orientation) to make real distance and speed measurements on the image and extrapolation of the sizes of the objects to be detected for each pixel using basic geometry. Preferably, one of those described in these procedures is used
20 the document by Hartley, R and Zisserman, A with the title “Multiple view geometry in computer vision” of Cambridge University Press 2003. - Weak calibration procedure in which from a series of observations of a type of object to be detected, the size of the object to be detected is associated to each pixel of the image either by brute force (1 observation / 1 pixel) or by
25 interpolation of an estimated geometric model based on a limited number of observations. Preferably, within this procedure the one described in patent application US7596241 or patent application ES2364915 is used.
In addition to these two types of calibration procedures, in a preferred embodiment
30 of the image improvement procedure for video analysis or automatic video surveillance systems to obtain the depth or scene information, the calibration procedure described in Spanish patent application P201330462 is used. This calibration procedure is a weak calibration in terms of parameter estimation (since it only refers to extrinsic ones) but robust in terms of accuracy
35 and correction of errors and includes, as explained in the aforementioned application, at least the following phases:

- Sample acquisition phase that is divided into the following sub phases:
or sub phase of image acquisition;
or sub phase of image processing through which it is determined whether there is any moving object in said image; Y
5 or sub-phase of pre-classification of persons that determines whether the mobile object identified is a person or not, the size and position of the identified mobile object being stored as a sample; Y
- Calibration phase that obtains the size of a person for each image position from the size and position data obtained for each object 10 identified as a person in the sample acquisition phase.
Once the depth or scene information of the image has been obtained, it is possible to go into detail the procedure for improving the image itself. For this, it should be noted that there are two major types of image enhancement procedures
15 called spatial filtering and equalization.
Accordingly, it is the object of the invention to apply this image depth or scene information to this type of image enhancement procedures in order to improve the contrast of the videoanalysis or automatic video surveillance images.
20 Regarding the first type, the so-called spatial filtering, it should be noted that it is based on selecting and applying a filter on the image with the aim of reducing noise, increasing the details, smoothing the image ..., obtaining an image with better contrast These filters are nothing more than small sub-images that are convolved with the main image and
25 generate an answer. Thus, depending on the size of the filter we will deal with different scales of image information.
Preferably, the spatial filtering and how to apply it is based on at least the following steps:
30 -For each point of the image the size in pixels of the object to be detected (, ℎ) is obtained by means of the calibration system, being � and ℎ respectively the width and height in pixels of the object to be detected at that point.
- For each point of the image a spatial filter of size between 3 × 3 pixels and max (, ℎ) × max (, ℎ) is constructed, and 35 Each point of the image is convolved with the spatial filter of the size corresponding to that point .

Well, it is the object of the present invention that said filters adjust their size in a variable manner at each image position based on the image or scene depth information obtained during the calibration phase. Preferably, said information should comprise at least the size of the object to be detected that was estimated during calibration.
Thus, for example, if a spatial filtering is used to reduce the noise, it is possible to choose between those filters that reduce the noise in the spatial domain, among which the linear filters would be found, such as: average filter, and those not linear, such as: medium or bilateral filter, and those filters that reduce noise in the transformed domain, among which filters based on the wavelet transform would be found. However, regardless of the filter that is chosen, it must adapt its size in each position of the image depending on the depth or scene information obtained during the calibration phase.
Also, in relation to other techniques such as details enhancement or high-pass filtering, they should adjust the sizes of their filters based on the depth or scene information obtained during the calibration phase.
With regard to the second type, that of equalization, it should be noted that the equalization of an image is actually the normalization of its histogram. That is, the main idea is that, given an input image, said input image is transformed based on the information of its histogram so that the output image has a histogram as flat as possible.
Thus, if we define � as the input image whose possible values � range from 0≤≤ − 1, the probability that the input image has the value is defined as:
() = (=) =
where � is the total number of pixels whose value is equal to � and � to the total number of pixels in the image. Through this histogram we have a reference of the probability of each gray level in the input image. The next step is to apply a transformation T to the entire input image I so that the output image = () has a astogram as flat as possible and thus improve the contrast of the image and, consequently, its quality. This output image � is also called an equalized image.

In a preferred embodiment, the transformation function is that indicated on page 91 of the book by González and Woods entitled "Digital Image Processing" of Prentice Hall 2002.
5 In short, this simple equalization comprises at least the following steps: 1-Calculate the histogram () of the entire input image and its corresponding transformation function � 2-Apply that transformation over the entire input image and thus obtain the equalized image 10 = ()
It should be noted that in general the input and output images are defined with the same number of bits, 8 bits per pixel, with a range of 256 possibilities (28). However, some acquisition devices, such as thermal cameras,
15 use 14 bits per pixel so keeping the output image with 8 bits per pixel equalization should reduce the space of 16,384 possibilities (214) of the input image to 256 possibilities of the output image.
Now, as with the image enhancement procedure called spatial filtering, the depth or scene information of the image must be used to significantly improve the contrast of the image.
In the case of equalization, depth or scene information is used to focus the improvement of the image in those areas in which the objects to be detected by the video analysis system are smaller, making it difficult to find in difficult areas Contrast detection is as large as possible.
That is, step number 1 of the procedure for simple equalization is modified by calculating the transformation T using only the information of the pixels of the 30 regions where the objects to be detected are smaller.
In this way, the contrast in difficult detection areas is maximized even if the easy detection zones (large objects to be detected) can be harmed.
35 To do this, you must also define the regions of the image in which the objects to be detected are smaller, which is called the region of interest (ROI).

This region is not restricted to any specific shape or size to the point that it could well be made up of subregions. Preferably, it is defined as a rectangle � = [,,, ℎ]
5 where� e correspond to the coordinates of the upper corner of, while
and ℎwith the width and height values of the region in pixels.
On the other hand, the sub-image formed by the pixels of the input image � contained in the region is defined as � and the histogram of this sub-image as
() = (� =) = � � 10 where � is the total number of pixels of the sub-image � whose value is equal to � and � atTotal number of pixels in the sub-image. So if in the simple equalization ofhistogram the transformation was obtained as a function of the histogram of the image of
input = (()), a new transformation is defined as the transformation function calculated based on the histogram of the pixels of the region of interest
� �
15 = (()). So, the image enhancement procedure based on equalization �
Simple comprises at least the following steps:
1-Calculate the histogram of the pixels of the input image contained in the region of interest () and use this information to obtain its transformation function

20 corresponding; 2-Apply that transformation � over the entire input image and thus obtain the equalized image = ()
In this way the contrast is maximized in hard to detect areas although
25 may damage some areas of easy detection (those in which the objects to be detected are larger). Brief description of the drawings
30 The foregoing and other advantages and features will be more fully understood from the following detailed description of some embodiments with reference to the attached drawings, which should be considered by way of illustration and not limitation, in which:

- fig. 1 illustrates the block diagram of a video analysis system or
video surveillance according to the invention; - fig. 2 shows the block diagram of the detection system; - fig. 3 represents the block diagram of a scene calibration system
5 based on a strong calibration procedure; - fig. 4 illustrates the block diagram of a scene calibration system based on a weak calibration procedure; - fig. 5 shows an image to which a scene calibration system has been applied; 10-fig. 6 represents an equalization procedure. - fig. 7 illustrates the operation of a hysteresis based switch. Detailed description of an embodiment example
Fig. 1 illustrates the block diagram of a video analysis or automatic surveillance system (1) according to the invention comprising at least one image acquisition device (2) from which images of an area of space are obtained, an image scanning system (3) that provides the digital image obtained by said image acquisition device (2), a system for processing the
20 image (4) and two alternative operating systems, the scene calibration system (5) and the detection system (6).
The image acquisition devices (2) allow to obtain images in the spectrumIR or thermal. Preferably they are fixed cameras with this type of image capture.
25 Also included are image acquisition devices (2) that obtain images in the near IR spectrum, such as day / night cameras operating with this section of the electromagnetic spectrum during night surveillance.
In this regard, it should be noted that certain image acquisition devices (2) already
30 provide a digital image, integrating them an image digitization system (3) so that it would not be necessary to include said image digitization system (3) in the video analysis or automatic surveillance system (1).
The image acquisition device (2) in the case that it already allows obtaining a digital image or the image scanning system (3) can be prepared to transmit the images by any means of transmission (cable, fiber, wireless, etc.). .).

The image processing system (4) applies at least one image improvement procedure such that at the exit of said system the image is of sufficient quality to detect a particular type of object, preferably, a
5 person
As mentioned, the video analysis or automatic surveillance system (1) has two alternative operating systems, the detection system (6) and the scene calibration system (5).
10 The detection system (6) is applied regularly during the operation of the video analysis or automatic surveillance system (1) since it is the one that allows the detection of specific objects, preferably people.
15 The scene calibration system (5) is preferably applied only once at the start of the start-up of the video analysis or automatic surveillance system (1) and must provide the image with a spatial reference so that The detection system (6) can reference all the calculations it performs during the detection process: calculation of the distance traveled, speed, size of the objects, etc., as well as providing also
20 direct or indirect information on the depth of the actual scene captured in the image for the image processing system (4).
Preferably, the scene calibration system (5) can be any type of system that obtains the depth of the actual scene captured in the image either directly
25 or indirect. In a preferred embodiment, the scene calibration system (5) is a system that obtains the variation of the approximate size of the object to be detected for each of the pixel coordinates of the image since it is an indirect way of measuring the depth of the actual scene captured in the image.
30 Fig. 2 shows the block diagram of the detection system (6) comprising a static scene segmentation system (7), a candidate generation system (8), a classification system (9), a monitoring system (10) and a decision system (20).
35 The static scene segmentation system (7) classifies the pixels into at least two types, moving objects and objects belonging to the background of the image.

The candidate generation system (8) groups the pixels that refer to moving objects and assigns a unique identifier to each moving object of the image.
5 It should be noted that both for the static segmentation system (7) and for thecandidate generation system (8) it is very relevant to have an imagesufficiently contrasted.
The classification system (9) classifies moving objects according to whether it is an object to be detected,
10 preferably, person and / or vehicle, or it is not. This system needs, as just mentioned, the scene information obtained during the calibration phase to perform the necessary calculations (speed measurement, size, distance traveled, etc.) to classify the objects and hence in the block diagram of Fig. 2 a block called calibration appears to reference such a need for information.
15 A tracking system (10) maintains the temporal coherence of the objects to finally, depending on the detection rules introduced by the user generate the respective intrusion alarms.
20 The decision system (20) is responsible for determining, based on some rules - hence there is a block called rules to reference such need for information - if the objects classified by the classification system (9) should be considered as intruders, generating the corresponding alarm in that case.
25 A block diagram of a scene calibration system is shown in Fig. 3
(5) based on a strong calibration procedure, such as that described in the Hartley, R and Zisserman document, A with the title "Multiple view geometry in computer vision" of Cambridge University Press 2003.
Said calibration system (5) may comprise at least one parameter insertion system of the image acquisition device (14) and a scene parameter calculation system (15).
35 The parameter insertion system of the image acquisition device (14) obtains, either directly or through the user itself, the intrinsic parameters of the image acquisition device.

image acquisition (2), such as: focal length, pixel size, radial lens distortion, and extrinsic parameters, such as: height and angular orientation.
The scene parameter calculation system (15) obtains the size of the objects to5 detect for each pixel.
A block diagram of a scene calibration system (5) based on a weak calibration procedure is illustrated in Fig. 4, such as those described in patent application US7596241 or patent application ES2364915.
10 As can be seen, said scene calibration system (5) comprises at least one static scene segmentation system (7), a candidate generation system (8), a tracking system (10), an observed size / position mapping system (11) and scene parameter estimation system (12).
Preferably, the static scene segmentation systems (7), candidate generation (8), tracking (10) perform the same functions as those described in the detection system (6), and may even be the same.
20 The size / position mapping system (11) obtains the variation of the approximate size of the object to be detected for each of the pixel coordinates of the image.
The scene parameter estimation system (12) allows obtaining other parameters necessary for the detection system (6), such as: speed measurement, size and distance traveled.
In a preferred embodiment, the scene calibration system (5) of the video analysis or automatic surveillance system (1) according to the invention uses the calibration procedure described in Spanish patent application P201330462 to obtain the
30 depth or scene information.
Regardless of the calibration procedure used by the scene calibration system (5), Fig. 5 shows an image to which the calibration system has been applied and in which the rectangles (16) indicate the approximate size of the object to
35 preferably detect people, at the point where the rectangle (16) is drawn.

The image processing system (4) according to the invention performs an image enhancement process comprising a processing step in which through said depth or scene information, entered by the user or obtained through any scene calibration system (5), although preferably those using the procedures just described, the contrast of the images captured by the image acquisition device (2) is improved.
In a preferred embodiment, the image processing system (4) comprises at least one filter that adjusts its size in a variable manner at each image position based on the image or scene depth information, preferably, a percentage of the size of the object to be detected that has been estimated by the scene calibration system (5).
In another preferred embodiment, the image processing system (4) comprises at least one equalization procedure in which depth information is used
or scene obtained through the scene calibration system (5).
Preferably, said equalization procedure is based on a simple equalization procedure centered on a region that is considered of interest (r). The criteria for defining said region of interest are preferably:
- User defined region manually; O well
- Region in which the objects to be detected are smaller; O well
- Transit areas during the scene calibration process (5), based on the Spanish patent application P201330462.
In the case of the criterion in which the objects to be detected are smaller, preferably, two person sizes are defined in pixels: � (minimum object size that is capable of detecting the detection system (6) and which is given by the detection system itself (6)) and which corresponds to the maximum possible size of an object to be detected in the image to be considered (which is given by the calibration system (5)). Thus, in this case, the criterion for defining the ROI will be all those pixels for which the expected object size given by the scene calibration system (5) is in the range (, + (-)). Being � a number between 0 and 1 that allows to regulate the level of equalization. As explained above, there are calibration procedures that consist of associating observed object sizes to the position where they have been

observed to, in this way, estimate the geometric model that describes the scene, among which, preferably, the scene calibration system (5) uses the one described in Spanish patent application P201330462. Preferably, in these cases, a criterion for defining the region of interest (r) comprises at least the following steps:
- Divide the image into N cells of adjustable size;
- Mark the cells in which the calibration system (5) has obtained at least one sample.
- Define the region of interest as the convex zone that surrounds the marked cells.
The regions of interest (r) are not restricted to any form or size.
referenceably, said region of interest (r) is defined as a rectangle (17)
� = [,,, ℎ] where� e correspond to the coordinates of the upper corner of, while
and with the width and height values of the region in pixels.
In video analysis or video surveillance scenes, the relevant content, at the level of object detection, is usually centered in the central part of the image, so it is preferably defined
� = [,,, ℎ] = [,,,]
where ,, ∈ (0,1) and (+) <1; � is the total width of the input image; and � is the vertical coordinate that delimits the detection limit (preferably, the vertical coordinate from which the expected size of the object to be detected by a person, is smaller than the minimum size that the system needs to be able to detect a person) that It can be entered by the user or it can be obtained from the calibration system (5).
In Fig. 5 a rectangle (17) is drawn that defines a region of interest for that image.
It should be noted that this type of rectangular region is suitable for any scene calibration system (5) since the final result of the calibration is a person size map per pixel.

In a preferred embodiment, the equalization procedure defines a sub-image formed by the pixels of the input image contained in the region � as � and the histogram of this sub-image as
() = (� =) = �
Likewise, a new transformation is defined as the transformation function
5 calculated based on the histogram of the pixels of the region of interest = (()).
�
In short, in a preferred embodiment of the equalization procedure it comprises at least the following steps:
10 1-Calculate the histogram of the pixels of the input image contained in the region
of interest () and use this information to obtain its transformation function

correspondent ; 2-Apply that transformation over the entire input image and thus obtain the equalized image = ()
15 As mentioned previously, the fact of equalizing certain areas of the image with a histogram that does not correspond to it can impair the contrast of these areas with the appearance of unwanted effects such as noise, saturation in the gray level , etc. Thus, in a preferred embodiment, a smoothing stage of the
20 equalization by a weighted sum of the image equalized with the previous method and the unqualified image as follows:
(,) = (,) · (,) + 1 - (,) · (,)

Where the weighting function (,) can be any type of function whose value in the
The center of the region of interest is maximum, although preferably it is a two-dimensional Gaussian function centered on the center of the region of interest (r) and with standard deviations of its dimensions the width and height of the region of interest itself (r), leaving the center of the Gaussian as:
=, = (+, � -)
30 and its standard deviation vector as

=, = (, �)Consequently, the function (,)
−1 − � 2−
(,) = (+ () 2)
2 � �
As can be seen, the value of (,) in the center of the region of interest (r) is
5 maximum (equal to 1) and, as the values of x or y move away from the center, the value of (,) decreases and therefore the unqualified image begins to take relevance given that the function 1 − g (x , and) grows.
Consequently, this entire stage performs a smoothing of the equalization being able to
10 also understand how the introduction of an artificial spotlight that illuminates the area of the region of interest (r).
So far, a basic equalization procedure has been described in which the depth or scene information obtained through the calibration system is used.
15 scene (5). However, two types of equalization procedure can be considered depending on the nature of the input image, local equalization procedure and remote equalization procedure.
The local equalization procedure is the simplest and is the one represented in Fig.
20 6. As can be seen, in this type of equalization procedure, the image of the image acquisition device (2) or the image resulting from applying the image digitization system (3) is used using depth information the one that has the scene calibration system (5) of the video analysis or automatic surveillance system (1).
25 Since the dynamic range of the image, that is, the range of values of most pixels, is sometimes too small, it may cause excessive noise to be introduced when expanding the histogram during equalization. This noise is not desirable in the image since it can lead to false alarms generated by the detection system
30 (6). For this reason, in a preferred embodiment, a step that studies the range of the histogram by calculating the entropy in the region of interest is incorporated into the equalization method according to the invention, a measure that, although indirect, is much more robust than Simply study the width of the histogram.

In this sense, the entropy of the image in the region of interest is defined as:
() = - () · log (())
�
This metric will be larger the more the histogram stores at a uniform probability distribution 5 and the wider the dynamic range of the image.
In this way, two threshold values are set for which we will activate (on) or deactivate (off) the equalization: � and.
10 The operation of this hysteresis-based switch is illustrated in Fig. 7. Specifically, if during the equalization procedure it is in the "off" equalization mode and the calculated entropy rises above, the activation procedure is activated. equalization On the contrary, if it is in the "on" state of equalization and the entropy falls below, the equalization procedure is deactivated.
15 It should be noted that this hysteresis cycle is implemented to avoid jumps between modes that may affect the detection system (6). However, and to further soften this transition between mode changes, we work with entropies calculated using a moving average instead of instantaneous entropies:() =(−1) · (1 -) + () · �
20 being a very small value between 0 and 1.
On the other hand, the remote equalization procedure is based on remotely defining the region of interest, based on the depth or scene information obtained in the image calibration system (5), for those image acquisition devices
(2) or image scanning systems (3) that have software that executes an equalization procedure. That is, the equalization procedure is performed by the image acquisition device (2) or the image scanning system (3) but on the region of interest defined from the depth or scene information
30 obtained in the image calibration system (5).

权利要求:
Claims (15)
[1]

one. Image enhancement procedure for video analysis or video surveillance systems
automatic comprising at least one image acquisition device
5 (2) through which an image is captured in the IR or thermal spectrum of an area
of space, which is digitized by said image acquisition device (2)
or by an image scanning system (3), a calibration system of
scene (5) and a detection system (6) through which at least one
object type
10
characterized because
said method comprises at least one processing stage in which
the contrast of the digital image captured by the acquisition device is improved
fifteen of images (2) through the depth or scene information of the image,
obtained, directly or indirectly, by the scene calibration system (5) or
well entered manually by the user and comprising at least one
equalization stage defined from depth or scene information
of the image obtained by the scene calibration system (5);
twenty
said equalization stage comprises at least the following steps
o Define a region of interest (r);
25 30 o Calculate the histogram of the pixels of the input image contained in said region of interest () and use this information to obtain a corresponding transformation function; where () = (� =) = � is the histogram of this region of interest formed by the image pixels contained in the region; and is a transformation function calculated based on the histogram of the pixels of the region of interest � = () � Apply that transformation over the entire input image and thus obtain the equalized image = ().

[2]
2. Image enhancement process for video analysis or automatic video surveillance systems according to claim 1 wherein the region of interest (r) is one in which the types of objects to be detected are smaller and its detection is difficult to detect. detection (6).
[3]
3. Image improvement procedure for video analysis or automatic video surveillance systems according to the preceding claim in which the region of interest (r) comprises at least all those pixels for which the expected size of the objects to be detected is in the range ( , + (-)); the minimum object size to be detected capable of detecting the detection system
(6); � the maximum possible size in the image of an object to be detected;
A number between 0 and 1.
[4]
4. Image enhancement procedure for video analysis or automatic video surveillance systems according to any of claims 2 or 3 wherein the region of interest is defined with a rectangle � = [,,, ℎ] = [,,,] where � and � correspond to the coordinates of the upper corner of, � and ℎ with the width and height values of the region in pixels, ,, ∈ (0,1) and (+) <1; � is the total width of the input image; and � is the vertical coordinate that delimits the limit of detection (from which the expected size of the object to be detected is smaller than the minimum size that the detection system (6) can detect.
[5]
5. Image enhancement procedure for video analysis or automatic video surveillance systems according to any one of claims 1 to 4 wherein the equalization stage comprises at least one step that estimates the range of the histogram of the region of interest by calculating entropy in said region of interest as follows:
() = - () · log (())
� and another step that sets at least two threshold values (� and) for which the equalization stage is activated or deactivated respectively.

[6]
6. Image improvement procedure for video analysis or automatic video surveillance systems according to the preceding claim in which the step in which the entropy is calculated is performed by the following moving average
() =(−1) · (1 -) + () · � 5 being a very small value between 0 and 1.
[7]
7. Image enhancement process for video analysis or automatic video surveillance systems according to any one of claims 1 to 6, further comprising a stage of equalization smoothing in which a new image is obtained at
10 from the weighted sum of the equalized image and the unqualified image as follows:
(,) = (,) · (,) + 1 - (,) · (,)
Where (,) is any function whose value is maximum and equal to 1 in the center of the region of interest (r); 15 I (x, y) is the unqualified digital image.
[8]
8. Image enhancement method for video analysis or automatic video surveillance systems according to claim 4 further comprising a step of
20 smoothing the equalization in which a new image is obtained from the weighted sum of the equalized image and the unqualified image as follows:
(,) = (,) · (,) + 1 - (,) · (,)
Where
−1 − � −� 2 � �
(,) = (2 + () 2);
=, � = +, � -;
25 =, = (, 2); 2
� �
I (x, y) is the unqualified digital image
[9]
9. Image enhancement procedure for automatic video analysis or video surveillance systems 30 comprising at least one image acquisition device

(2) through which an image is captured in the IR or thermal spectrum of an area of space, which is digitized by said image acquisition device (2)
or by an image scanning system (3), a calibration system of
scene (5) and a detection system (6) through which at least one5 object type
characterized because
said method comprises at least one processing stage in which
10 the contrast of the digital image captured by the image acquisition device (2) through the depth or scene information of the image, obtained, directly or indirectly, by the scene calibration system (10) is improved. 5) or manually entered by the user and comprising at least one stage of spatial filtering defined from the depth information or
15 scene of the image obtained by the scene calibration system (5);
said spatial filtering stage comprises at least the following steps:
o For each point of the image you get the size in pixels of the
20 object to be detected (, ℎ) by means of the scene calibration system (5), being � and ℎ respectively the width and height in pixels of the object to be detected at that point.
o For each point of the image a spatial filter of size between 3 × 3 pixels and max (, ℎ) × max (, ℎ) is constructed,
25 o Each point of the image is convolved with the spatial filter of the size corresponding to that point.
[10]
10. Image enhancement procedure for video analysis or automatic video surveillance systems according to any of the preceding claims in which the
Depth or scene information is obtained from a scene calibration system (5) that applies a calibration procedure comprising at least the following phases:
o Sample acquisition phase that is divided into the following sub phases:
or sub phase of image acquisition;
35 or sub phase of image processing through which it is determined whether there is any moving object in said image; Y

or sub-phase of pre-classification of persons that determines whether the mobile object identified is a person or not, the size and position data of the identified mobile object being stored as a sample; Y
5 o Calibration phase that obtains the size of a person for each position ofthe image from the size and position data obtained for each objectidentified as a person in the sample acquisition phase.
[11]
11. Image enhancement procedure for video analysis or video surveillance systems
Automatic according to the preceding claim when it depends on claims 1 to 8 in which the region of interest (r) is defined as the transit zone during the scene calibration procedure defined in the preceding claim.
[12]
12. Image enhancement procedure for video analysis or video surveillance systems
Automatic according to claim 10 wherein the region of interest (r) is obtained according to the method:
o Divide the image into N cells of adjustable size;
o Mark the cells in the scene calibration system (5)
obtained at least one sample in the sample acquisition phase 20 described in claim 10;
o Define the region of interest as the convex zone that surrounds the marked cells.
[13]
13. Image enhancement system for video analysis or video surveillance systems
25 for executing any of the image enhancement method according to the preceding claims wherein said image enhancement system comprises at least one image acquisition device (2) through which an image is captured in the IR spectrum or thermal of an area of space, which is digitized by said image acquisition device (2) or by a
30 image scanning system (3), a scene calibration system (5) and a detection system (6) through which at least one type of object is detected and comprising at least
a static segmentation system (7) that classifies the pixels into at least two types, moving objects and objects belonging to the background of the image;

a candidate generation system (8) that groups the related mobile pixels into objects;
a classification system (9) that classifies objects according to whether it is an object type5 to detect or not; Y
a tracking system (10) that maintains the temporal coherence of the objects to finally, depending on the detection rules introduced by the user, generate the respective intrusion alarms.
[14]
14. Image enhancement system for video analysis or automatic video surveillance systems according to the preceding claim wherein the image acquisition device (2) is analog and the digitalization of the image captured by said image acquisition device (2) It is prior to the processing stage.
[15]
15. Automatic video surveillance or video surveillance system characterized in that it has functional elements suitable for performing any of the image enhancement procedures of claims 1 to 12.

FIG. one
8 920 7
Tracking systemCalibrationRules
FIG. 2
FIG. 3

8 11 5
7
Tracking system
10
FIG. 4

FIG. 6
HHL HLH
FIG. 7

类似技术:

公开号 | 公开日 | 专利标题

CN109076198B|2021-01-29|Video-based object tracking occlusion detection system, method and equipment

WO2018052547A1|2018-03-22|An automatic scene calibration method for video analytics

US20150055824A1|2015-02-26|Method of detecting a main subject in an image

JP6125188B2|2017-05-10|Video processing method and apparatus

US20180240251A1|2018-08-23|Imaging system, object detection device and method of operating same

JP2015528614A5|2017-05-25|

JP5832910B2|2015-12-16|Image monitoring device

ES2563098B1|2016-11-29|IR image enhancement procedure based on scene information for video analysis

Bosch-Jorge et al.2014|Fall detection based on the gravity vector using a wide-angle camera

JP2009025910A|2009-02-05|Obstacle detection device, obstacle detection system, and obstacle detection method

Berger et al.2010|Room occupancy measurement using low-resolution infrared cameras

CN107145820B|2020-11-17|Binocular positioning method based on HOG characteristics and FAST algorithm

KR101769741B1|2017-08-21|Method and apparatus for recognizing iris through pupil detection

US10748294B2|2020-08-18|Method, system, and computer-readable recording medium for image object tracking

KR101341243B1|2013-12-12|Apparatus and method of restoring image damaged by weather condition

JP2013228915A|2013-11-07|Pedestrian detecting device and method for detecting pedestrian

Sheng et al.2016|Dark channel prior-based altitude extraction method for a single mountain remote sensing image

CN111275045A|2020-06-12|Method and device for identifying image subject, electronic equipment and medium

CN110020572B|2021-08-10|People counting method, device and equipment based on video image and storage medium

KR101958927B1|2019-03-15|Method And Apparatus for Providing Adaptive Counting People

ES2452790B1|2015-01-20|Procedure and image analysis system

Gundawar et al.2014|Improved single image dehazing by fusion

Zaihidee et al.2015|Comparison of human segmentation using thermal and color image in outdoor environment

Hoo et al.2015|Skin-based privacy filter for surveillance systems

Jo et al.2013|Performance improvement of human detection using thermal imaging cameras based on mahalanobis distance and edge orientation histogram

同族专利:

公开号 | 公开日

CA2989188A1|2016-12-22|

US10452922B2|2019-10-22|

GB201721740D0|2018-02-07|

WO2016203078A4|2017-09-14|

IL256202A|2021-05-31|

GB2557035A|2018-06-13|

US20180225522A1|2018-08-09|

WO2016203078A2|2016-12-22|

ES2563098B1|2016-11-29|

IL256202D0|2018-02-28|

GB2557035B|2021-05-26|

WO2016203078A3|2017-07-27|

引用文献:

公开号 | 申请日 | 公开日 | 申请人 | 专利标题

US5046118A|1990-02-06|1991-09-03|Eastman Kodak Company|Tone-scale generation method and apparatus for digital x-ray images|

WO1996016534A2|1994-11-25|1996-06-06|Sophisview Technologies, Ltd.|System and method for diagnosis of living tissue diseases|

US7308126B2|1997-08-28|2007-12-11|Icad, Inc.|Use of computer-aided detection system outputs in clinical practice|

US6696945B1|2001-10-09|2004-02-24|Diamondback Vision, Inc.|Video tripwire|

US7706576B1|2004-12-28|2010-04-27|Avaya Inc.|Dynamic video equalization of images using face-tracking|

US7596241B2|2005-06-30|2009-09-29|General Electric Company|System and method for automatic person counting and detection of specific events|

US8139828B2|2005-10-21|2012-03-20|Carestream Health, Inc.|Method for enhanced visualization of medical images|

US7826666B2|2008-02-27|2010-11-02|Honeywell International Inc.|Methods and apparatus for runway segmentation using sensor analysis|

US8059911B2|2008-09-30|2011-11-15|Himax Technologies Limited|Depth-based image enhancement|

US8339475B2|2008-12-19|2012-12-25|Qualcomm Incorporated|High dynamic range image combining|

US8274565B2|2008-12-31|2012-09-25|Iscon Video Imaging, Inc.|Systems and methods for concealed object detection|

US8054290B2|2009-05-27|2011-11-08|Microsoft Corporation|Image contrast enhancement in depth sensor|

SE536510C2|2012-02-21|2014-01-14|Flir Systems Ab|Imaging method for detail enhancement and noise reduction|

ES2452790B1|2013-03-28|2015-01-20|Davantis Technologies Sl|Procedure and image analysis system|

CN103400351B|2013-07-30|2015-12-23|武汉大学|Low light based on KINECT depth map shines image enchancing method and system|

US9460499B2|2014-05-30|2016-10-04|Shenzhen Mindray Bio-Medical Electronics Co., Ltd.|Systems and methods for selective enhancement of a region of interest in an image|KR20180040286A|2016-10-12|2018-04-20|삼성전자주식회사|Display apparatus and method of controlling thereof|

CN109658515A|2017-10-11|2019-04-19|阿里巴巴集团控股有限公司|Point cloud gridding method, device, equipment and computer storage medium|

CN109409345B|2018-12-24|2020-10-02|台州和沃文化传播有限公司|Intelligent playing performance device|

法律状态:
2016-11-29| FG2A| Definitive protection|Ref document number: 2563098 Country of ref document: ES Kind code of ref document: B1 Effective date: 20161129 |

优先权:

申请号 | 申请日 | 专利标题

ES201530836A|ES2563098B1|2015-06-15|2015-06-15|IR image enhancement procedure based on scene information for video analysis|ES201530836A| ES2563098B1|2015-06-15|2015-06-15|IR image enhancement procedure based on scene information for video analysis|

CA2989188A| CA2989188A1|2015-06-15|2016-06-13|Method for ir or thermal image enchancement based on scene information for video analysis|

PCT/ES2016/070443| WO2016203078A2|2015-06-15|2016-06-13|Ir or thermal image enhancement method based on background information for video analysis|

US15/580,257| US10452922B2|2015-06-15|2016-06-13|IR or thermal image enhancement method based on background information for video analysis|

GB1721740.7A| GB2557035B|2015-06-15|2016-06-13|Method for IR or thermal image enhancement based on scene information for video analysis|

IL256202A| IL256202A|2015-06-15|2017-12-08|Ir or thermal image enhancement method based on background information for video analysis|

[返回顶部]