WO2018130016A1

WO2018130016A1 - Parking detection method and device based on monitoring video

Info

Publication number: WO2018130016A1
Application number: PCT/CN2017/113067
Authority: WO
Inventors: 王鸿鹏; 尤磊; 牟蕾; 何华门; 柯宇
Original assignee: 哈尔滨工业大学深圳研究生院
Priority date: 2017-01-10
Filing date: 2017-11-27
Publication date: 2018-07-19
Also published as: CN106878674A; CN106878674B

Abstract

The present invention provides a parking detection method on the basis of monitoring video, with strong robustness. The method comprises: performing static target detection on the basis of a foreground history pixel and image similarity. The method mainly comprises steps of: extracting a foreground moving target through Gaussian Mixed Model; obtaining a suspicious stationary pixel area through a suspicious static pixel matrix; and calculating the image similarity. In a vehicle identification stage, an Haar classifier is improved, which avoids training process stagnation during the training of a cascaded strong classifier. It ensures that the stagnation phenomenon is avoided during the training process such that a stronger robustness in the training of the cascaded strong classifier is achieved. In a vehicle detection process, only a static target area and a neighboring area thereof detected through the static target detection are placed in the Haar classifier for detection, instead of putting the entire image into the Haar classifier. Thus, the amount of computation is greatly reducing, and the real-time performance of the algorithm is improved. In addition, an occlusion detection based on the Gaussian Mixture Model is also used to solve the problem of temporary occlusion, so as to reduce the missing rate of the algorithm.

Description

Parking detection method and device based on monitoring video

Technical field

The present invention relates to the field of computer vision and intelligent monitoring technologies, and in particular, to a parking detection method and apparatus based on surveillance video.

Background technique

With the rapid development of the national economy and the rapid increase of motor vehicles, the problem of urban transportation in China has become increasingly severe. In the past, parking detection mainly relied on manual monitoring and collection methods of fixed-point information. Greatly consumed the human, financial and material resources of the traffic control department. In recent years, with the rapid development of technology, the price of surveillance equipment such as video capture cards and cameras has gradually declined. The indoor and outdoor video surveillance systems have been widely applied to various occasions, such as indoor and outdoor parking lots, subways, banks, and highways. , hotels, supermarkets, campus buildings, etc. It plays an increasingly important role in the public safety of society. However, most of the domestic video surveillance systems are still in the traditional mode, playing the role of “record only and not judge”. Therefore, it is only possible to investigate the situation and evidence collection through video playback afterwards. There are defects in real-time forensics and alarms, and staff members are required to constantly monitor the activities in the scene. The shifts are on duty day and night, the workload is huge, and people are vulnerable to human fatigue. The impact of the misdetection and missed detection, which makes the difficulty of forensics increase, loses the significance of real-time monitoring of the monitoring system. On the other hand, as the scale of surveillance systems continues to expand, the amount of video data increases, and it becomes more and more difficult to obtain useful intelligence or information, and the search efficiency is low, and it is difficult to meet the needs of the monitoring system.

Parking detection is an important branch of vehicle detection. Vehicle detection technology is mostly used in outdoor environments where one or more high-definition cameras are used to monitor parking conditions on a particular street. If you can use the parking detection technology to accurately and timely detect illegal parking, you can effectively obtain evidence and alarm, effectively avoid further traffic jams and traffic accidents, thus maintaining the safety of people's lives and property.

For the problem of parking detection, experts and scholars at home and abroad have proposed a variety of methods based on parking detection, including:

First, the use of optical flow method for vehicle detection, due to the high computational complexity of the optical flow method, coupled with the current image size is large, resulting in poor real-time optical flow method.

Second, the vehicle features are extracted in the restricted area, and classified by SVM (support vector machine), so that the information of the vehicle is extracted from the picture.

Third, the use of Bayesian network for vehicle detection. The characteristics of the vehicle are first extracted, and the color of the car and the color of the non-vehicle are effectively distinguished by one color. Second, extract the color information and border information around the car. Interest, through the canny edge detection method that automatically adjusts the threshold. Finally, a Bayesian network is used for classification.

Fourth, use the Gaussian mixture model to separate the foreground from the background, extract the foreground information, determine whether there is parking in the target area, and update the background in real time.

Of course, there are still some problems to be solved in these commonly used methods:

1) Real-time, the current surveillance video is HD surveillance video, the resolution is in the millions and above, although the computer hardware level has reached a very high level, but it is more difficult to deal with massive data like surveillance video. of. How to use surveillance video to process parking in a timely and effective manner is one of the urgent problems to be solved.

2) Robustness, the scenes of surveillance video are mostly in complex scenes, which cause great interference to parking detection. In complex scenes, there are many objects similar to vehicles, which brings great detection to vehicles. The problem. At present, most parking detection algorithms only identify vehicles by simple contour features, duty cycles or area features, resulting in higher false detection rates. In addition to the above-mentioned interference, there are interferences of illumination changes, interferences of occlusion, and the like.

3) The practicability of the algorithm, when the parking detection algorithm has good real-time and robustness, it often requires a large amount of hardware resources. For real-life surveillance cameras, the processor is often not very high. If you want to achieve better real-time and robustness, it is relatively difficult to reduce the computational complexity of the algorithm and improve the computational efficiency. .

Summary of the invention

The object of the present invention is to overcome the problems in the prior art, and to provide a parking detection method and device based on monitoring video, which can automatically identify vehicles that are stopped within the coverage of the surveillance video without requiring much manual intervention. This kind of solution can solve the shortcomings of high cost, easy to be affected by human factors and waste of manpower in the overall relying on manual parking detection.

To achieve the above object, the present invention is achieved by the following technical solutions:

A parking detection method based on surveillance video includes the following steps:

Video data acquisition step: the video data captured by the surveillance camera of the ordinary roadside, the relative position of the surveillance camera and the angle of the shooting remain unchanged;

Background model establishing step: establishing a background model based on a mixed Gaussian model by analyzing a video sequence;

a foreground moving target extraction step: extracting a moving foreground target in the video image;

The stationary target area detecting step: detecting the still area in the video frame, specifically finding the suspicious still target area first, and then after N frames, comparing the similarity with the corresponding position of the previous frame image to determine whether it is a true still area. ;

Vehicle identification step: the improved Haar classifier is used to identify the stationary area to achieve the purpose of detecting the vehicle; by training a plurality of strong classifiers, the trained strong classifiers are cascaded to form the improved Haar classifier, one by one. The input image is used for vehicle identification, and after being identified, it is input to the next-level classifier, and if it is not passed, it is directly recognized as a non-vehicle. #

The invention has the beneficial effects that the vehicle stopped in the range covered by the surveillance video is detected by the surveillance video on the outdoor road, and the solution can solve the high cost and the human factor which are comprehensive in the manual parking detection. Impact, waste of manpower and other shortcomings.

DRAWINGS

Figure 1 is a flow chart of the method of the present invention;

Figure 2 is a flow chart of foreground detection employed by the present invention;

3 is a flow chart of image similarity calculation;

4 is a schematic diagram of the structure of a modified Haar classifier employed in the present invention.

Specific implementation

The present invention will be further described in detail below with reference to the accompanying drawings.

As shown in FIG. 1, the monitoring video based parking detection method of the present invention is mainly divided into several steps:

Video data acquisition step: The video data is taken as the video data of the ordinary roadside surveillance camera. The captured area is a long road, and the relative position of the surveillance camera and the angle of the shooting remain unchanged.

The background model establishing step: selecting the first N frames of video sequences from the video sequence to establish an adaptive mixed Gaussian background model, and continuously updating the background to adapt the model to changes in external conditions.

Foreground motion target extraction and shadow detection steps: From the X frame of the video, by comparing the background with the background, extracting the foreground moving target of each frame and using the shadow detection method, most of the shadows can be eliminated.

The still pixel matrix establishing step: after obtaining the foreground, a suspected still pixel matrix is established, and the size of the suspected still pixel matrix is the same as the relative position of each frame in the video, and is initialized to 0.

Suspicious quiescent area detection step: When a point is a foreground pixel, the value of the relative position of the suspected still pixel matrix is incremented by one. If a point has a foreground into the background, the value of the point in the still pixel matrix is set to zero. If a point reaches the threshold set as expected, the point is considered to be a suspected still pixel, and the connected area consisting of suspected still pixels is a suspected still area, and then the area is The information is stored in the picture buffer so that the next step is to confirm whether it is a real still area.

Resting area detecting step: after N frames, the corresponding picture information of the buffer is taken out, and then the degree of "the same" of the corresponding positions of the two pictures is determined (usually about 1500, about one minute in time, in It is assumed here that the external environment has not changed during this short period of time. When the "same" level is high, the suspicious stationary target is considered to be a true stationary target.

Vehicle identification step: Finally, the Haar classifier is used to identify the vehicle in the stationary area and its neighbors to determine whether it is a vehicle. The purpose of vehicle detection is achieved, and the occlusion detection based on the mixed Gaussian model is adopted to reduce the system miss detection rate.

In the technical solution of the present invention:

In the video data acquisition step: the video data is taken as the video data of the ordinary roadside surveillance camera, the captured area is a long road, and the relative position of the surveillance video head and the angle of the shooting remain unchanged, representative The public data set includes the i-LIDS parking detection data set and the PETS vehicle tracking data set. The i-LIDS parking detection dataset consists of approximately 50,000 consecutive surveillance video images, including four scenes: simple scene, medium scene, difficult scene and mixed scene. Among them, the simple scene is that there is little occlusion, the weather is better and the vehicle is stopped closer to the surveillance video, and the medium scene has more partial occlusion. The weather is shaking in the windy weather camera and stopping the vehicle far from the surveillance video, it is difficult The scene is more occluded, there is windy weather and the distance from the camera is far away. The weather conditions of the mixed scene change greatly, there are four parking phenomena and there are more partial occlusions. The PETS dataset includes a total of approximately 10,000 frames, including four parking lots, with good illumination and partial occlusion.

In the background model establishing step: the first N frames of video sequences are selected from the video sequence to establish an adaptive mixed Gaussian background model, and the background is continuously updated to adapt the model to changes in external conditions. Because, in the case of fixed cameras, the background changes are slow, and most of them are the effects of light, wind, etc. By modeling the background, separating the foreground and background from a given image, in general, the foreground is Moving objects to achieve the purpose of moving objects. The mixed Gaussian model uses K single Gaussian models to characterize the individual pixels in the image. The new Gaussian model is updated after the new frame is acquired, and each pixel in the current image is matched with the mixed Gaussian model. If successful, the decision is made. This point is the background point, otherwise it is the former attraction. The mathematical expression is as follows:

For a mixed Gaussian model with K Gaussian components, the probability that a pixel is x _N at time N is as follows:

Where w _j is the weight of the jth model, and η(x _N ; θ _j ) is the normal distribution of the jth Gaussian model expressed as

formula:

Where μ _k represents the mean of the K-gaussian model,

I represents the covariance of the K-gaussian model.

In the mixed Gaussian model, the background is the most important part of the model, and the value of the covariance is relatively small. For this reason, we sort the K Gaussian functions in descending order of w _k /σ _k , ranking first in the sequence. The Gaussian distribution is most likely to describe the Gaussian function of the current background. Through the above analysis, the first H Gaussian distributions are selected as the current background model, where H is obtained by:

Where T is the threshold.

In the foreground moving target extraction and shadow detection steps:

1) foreground detection, based on the previously established mixed Gaussian model, assuming that the current pixel value is x _i,t , and matching it with the K Gaussian models in order of priority, if the current pixel value is equal to the mean of a Gaussian model The absolute value is less than 2.5 times the variance, as shown in the following equation:

If the pixel is considered to belong to this Gaussian model, the point is considered to be the background. If the pixel does not match any of the Gaussian models, then its foreground is considered to be the moving target of the foreground. The detection process is shown in Figure 2.

2) Moving target shadow detection, the extracted foreground moving target may still contain the shadow of the moving target, in order to reduce the influence of the shadow on the foreground moving target extraction, the shadow must be removed. In order to ensure the real-time performance of the system, a shadow detection method based on HSV space is adopted. Refer to the following formula to determine whether a moving point is a shadow:

Wherein D and B respectively represent pixel points of the foreground image and pixel points of the background image in the HSV color space, wherein D(x, y).V, B(x, y).V respectively represent the point (x, y) The brightness value of the corresponding front spot and the brightness value of the background pixel. D(x, y).H, B(x, y).H represent the chrominance values of the foreground and background pixels of the corresponding point, respectively. D(x, y).S, B(x, y).S respectively represent the saturation values of the foreground and background pixels of the corresponding point. α and β represent the threshold of luminance, τ _s represents the threshold of saturation, and τ _H represents the threshold of chromaticity. The result is represented by S. If the result is 1 indicating that the point is a shaded pixel, if the result is 0, the point is a non-shaded point. The method has the advantages of strong versatility, fast calculation speed and high accuracy.

In the stationary target detection, it is mainly divided into two parts. The first part is the detection of suspicious static target area, and the second part is the picture similarity calculation. The main purpose is to detect whether the suspicious static target changes, so as to determine whether it is truly static. aims.

The step of establishing the still pixel matrix is specifically: the short-term time when the current scene target changes from the motion state to the stationary state, the still object is still the foreground, and whether the pixel point in the short time is the foreground is utilized. The information establishes a suspected still pixel matrix. The size of the still pixel matrix is the same as the size of the picture. At the beginning of the still pixel matrix, each element is initialized to 0, and B(x, y) represents whether the point at the position (x, y) of the current frame is foreground or not.

The method of updating the still pixel matrix is as follows

Where F _n (x, y) represents the value of the nth frame at which the still pixel matrix point (x, y) is located.

The step of detecting the suspected still region is specifically: when a certain point is a foreground pixel, the value of the relative position of the suspected still pixel matrix is incremented by one. If a point has a foreground into the background, the value of the point in the still pixel matrix is set to zero. If a point reaches the threshold that is expected to be set, and the value set here is 150, the point is considered to be a suspicious still pixel, and the connected area consisting of suspected still pixels is a suspicious still area, and then the information of the area is stored. Image buffer so that the next step is to confirm if it is a real still area.

The step of detecting the still area is specifically: after the suspicious still area is obtained, the information of the suspect still area is stored in the image buffer, and it is necessary to confirm whether it is a true still area after passing N frames. Whether the previous image and the image content after the N frame are "same" can be confirmed by the image similarity. As shown in Figure 3, the specific steps of image similarity calculation are as follows:

(1) Two images are read and grayscaled, and are labeled as image A and image B, respectively.

(2) Calculate the gray histograms of the two images, respectively labeled A_hist and B_hist.

(3) Normalize A_hist and B_hist, and the normalized formula is as follows:

x=(x-Min)/(Max-Min)

Where x represents the value corresponding to each pixel level in the histogram, Max represents the maximum value in the histogram, and Min represents the minimum value in the histogram.

(4) For the A_hist and B_hist after normalization, find the Pap s distance between them. The Barthel's formula is as follows:

Where DB(A_hist, B_hist) represents the Pap sm distance between A_hist and B_hist

(5) Determine the Barthel distance between the two gray histograms. If the Pap smear distance is too large, the degree of similarity between the two images is considered to be low. If the Pap sm distance is small, the similarity of the two images is considered to be higher.

(6) When judging that the Paging distance of the two images is small, the degree of difference in the relative positions of the two images is continuously judged. Reduce the two images to 32×32. Reduce the amount of calculations and reduce the impact of detail.

(7) Calculate the Euclidean distance of the two images after reduction, and the formula is as follows:

Where D(A, B) is the Euclidean distance after the image A and the image B are reduced to 32×32. When D(A, B) is too large, it can be judged that the contents of the two images are very different.

Through the above steps, it can be well solved whether the image content of the relative positions of the two images separated by a short time changes when the camera is not moving. It is thus possible to use this method to detect whether an object that has temporarily stopped moving has actually stopped. And it can solve the problem of occlusion of a small part of the target and improve the robustness of the algorithm.

The vehicle identification step is specifically: firstly, the improvement of the Haar classifier is implemented, a cascade classifier is designed, and finally the improved classifier is used to identify the stationary vehicle, and the occlusion detection is added, and the response is complicated. Unrecognized issues in the scene.

1) Cascade classifier, the main goal is to have a higher recognition rate and a smaller misrecognition rate. By training multiple strong classifiers, the trained strong classifiers are cascaded. The specific flow chart is shown in Figure 4, and the algorithm is as follows:

Through the improved algorithm above, the fault tolerance and robustness of the cascaded classifier training process can be improved, and the accuracy and false detection rate of the recognition are ensured, and the hysteresis phenomenon and the training time are avoided during the training process. The phenomenon.

2) Identification of the stationary vehicle, using the sub-image of the current stationary target area and its corresponding domain as the input of the Haar classifier, instead of taking the entire image as input, in order to reduce the Haar classifier in the vehicle When detecting, the search range of the detection window is reduced, and the real-time performance of the system is increased. The specific algorithm is as follows:

In the occlusion detection, the hybrid Gaussian model-based method can solve the problem that the recognition target is temporarily occluded in a complicated scene. When it is judged whether the stationary target is a car, since the occlusion often leads to the recognition failure, the mixed Gaussian model has a good adaptability to the external condition change, and the foreground of the motion in the current image can be accurately extracted. The specific occlusion detection steps are as follows:

a. Record the position rectangular frame information of the previous frame of the stationary vehicle.

b. Calculate the relative position and proportion of the foreground moving target pixel of the current frame in the current rectangular frame.

c. If the proportion is greater than 30%, it is determined that the stationary target has been operated.

d. If the proportion is less than or equal to 30%, it is determined that the stationary target is not running, and is still a stationary target, and the current rectangular information frame is used as the position information of the stationary vehicle in the current frame, and the alarm is continued.

The above is a further detailed description of the present invention in connection with the specific preferred embodiments, and the specific embodiments of the present invention are not limited to the description. It will be apparent to those skilled in the art that the present invention may be made without departing from the spirit and scope of the invention.

Claims

A parking detection method based on surveillance video, characterized in that the method comprises the following steps:

Video data acquisition step: the video data captured by the surveillance camera of the ordinary roadside, the relative position of the surveillance camera and the angle of the shooting remain unchanged;

Background model establishing step: establishing a background model based on a mixed Gaussian model by analyzing a video sequence;

a foreground moving target extraction step: extracting a moving foreground target in the video image;

The stationary target area detecting step: detecting the static area in the video, specifically finding the suspicious still target area first, and then after N frames, comparing the similarity with the corresponding position of the previous frame image to determine whether it is a true still area;

Vehicle identification step: the improved Haar classifier is used to identify the stationary area to achieve the purpose of detecting the vehicle; by training a plurality of strong classifiers, the trained strong classifiers are cascaded to form the improved Haar classifier, one by one. The input image is classified and identified, and after being identified, it is input to the next-level classifier, and if it is not passed, it is directly recognized as a non-vehicle.
The method according to claim 1, wherein the method further comprises a shadow detecting step of: reducing the influence of the shadow on the foreground moving target extraction by using a shadow detection method based on the HSV space, specifically: determining whether a moving point is The shadow refers to the following formula:

Wherein D and B respectively represent pixel points of the foreground image and pixel points of the background image in the HSV color space, wherein D(x, y).V, B(x, y).V respectively represent the point (x, y) The brightness value of the corresponding front sight and the brightness value of the background pixel; D(x, y).H, B(x, y).H respectively represent the chrominance value of the foreground pixel and the background pixel of the corresponding point; D(x , y).S, B(x, y).S respectively represent the saturation values of the foreground and background pixels of the corresponding point; α and β represent the threshold of brightness, τ s represents the threshold of saturation, and τ H represents the chromaticity Threshold; the result is represented by S. If the result is 1 indicating that the point is a shaded pixel, if the result is 0, the point is a non-shaded point.
The method according to claim 1, wherein said stationary target region detecting step comprises a suspicious still region detecting substep: establishing a still pixel matrix, and when a certain point is a foreground pixel, the relative position of the suspected still pixel matrix The value of the point will be increased by one; if a point has a foreground into the background, the point is 0 in the still pixel matrix; if a point reaches the threshold set by the expected value, the point is considered to be a suspicious still pixel, The connected area consisting of suspicious still pixels is suspicious Stop the area.
The method according to claim 3, characterized in that before the suspected still region detecting step, the method further comprises the establishment of a still pixel matrix: a short time period in which the current scene target changes from a motion state to a stationary state Inside, the still object is still foreground, and the suspected still pixel matrix is established by using the information of the pixel in the short time period for the foreground information. The size of the still pixel matrix is the same as the size of the picture, and the static pixel matrix initializes each element. 0, B(x, y) represents whether the point at the position (x, y) of the current frame is foreground, which is expressed as follows:

The method of updating the still pixel matrix is as follows:

Where F n (x, y) represents the value of the nth frame at which the still pixel matrix point (x, y) is located.
The method according to claim 1, wherein in the background model establishing step, the background is mainly modeled to facilitate separation of the foreground and background of the image, and the mixed Gaussian model is used by K single Gaussian. The model is used to represent the features of each pixel in the image, and the mixed Gaussian model is updated after the new frame image is obtained, and each pixel in the current image is matched with the mixed Gaussian model. If successful, the point is determined as the background point, otherwise Former sights.
The method according to claim 3, wherein in the still area detecting step, the suspect still area is verified, and the true still area is screened by comparing with the image after the Nth frame; as follows:

(1) reading two images and performing grayscale, respectively labeled as image A and image B;

(2) Calculate the gray histograms of the two images, respectively labeled A_hist and B_hist;

(3) Normalize A_hist and B_hist, and the normalized formula is as follows:

x=(x-Min)/(Max-Min)

Where x represents the value corresponding to each pixel level in the histogram, Max represents the maximum value in the histogram, and Min represents the minimum value in the histogram;

(4) For the A_hist and B_hist after normalization, find the Pap s distance between them, and the Pap s distance DB (A_hist, B_hist) is as follows:

(5) Determine the Barthel distance between the two gray histograms. If the Pap smear distance is too large, the degree of similarity between the two images is considered to be low. If the Pap address distance is small, the similarity of the two images is considered Higher degree;

(6) When judging that the Paging distance of the two images is small, continue to judge the degree of difference between the relative positions of the two images, reduce the two images to 32×32, reduce the calculation amount and reduce the influence of the details;

(7) Calculate the Euclidean distance between the two images after reduction

Where D(A, B) is the Euclidean distance after the image A and the image B are reduced to 32×32; when D(A, B) is too large, it can be determined that the content of the two images is very different.
The method according to claim 1, wherein in said vehicle identifying step, the classifier is applied to identify the stationary vehicle to determine whether it is a vehicle; to achieve the purpose of vehicle detection, and to adopt a hybrid Gaussian model Perform occlusion detection to reduce system miss detection rate; the specific algorithm is:

(1) acquiring a connected area of the still area and a length H and a width W of the picture;

(2) Calculating the coordinates (x, y) of the center point of the connected region;

(3) calculating a circumscribed rectangle R_static of the connected region of the still region;

(4) construct a rectangle R centered on the coordinate point (x, y), having a width of W/3 and a height of H/3;

(5) obtaining a sub-image img in a position where the rectangle R is covered in the original image;

(6) performing vehicle identification on the sub-picture img to obtain a recognition result enclosing frame;

(7) judging whether there is a intersection with R_static greater than 75% in the bounding box;

(8) If there is an alarm, there is no further detection.
The method according to claim 7, wherein in the step of recognizing the vehicle, the shadow detection is mainly to solve the problem that the target is temporarily occluded in a complicated scene, and the specific steps are as follows:

a. Record the position rectangular frame information of the previous frame of the stationary vehicle;

b. calculating the relative position and proportion of the foreground moving target pixel of the current frame in the current rectangular frame;

c. If the proportion is greater than 30%, it is determined that the stationary target has been operated;

d. If the proportion is less than or equal to 30%, it is determined that the stationary target is not running, and is still a stationary target, and the current rectangular information frame is used as the position information of the stationary vehicle in the current frame, and the alarm is continued.
A parking detection device based on surveillance video, characterized in that the device comprises:

Video data acquisition unit: for video data captured by a surveillance camera of an ordinary roadside, the relative position of the surveillance camera and the angle of the shooting remain unchanged;

Background model building unit: used to establish a background model based on mixed Gaussian model by analyzing video sequences type;

a foreground moving target extraction unit: for extracting a moving foreground target in the video image;

a stationary target area detecting unit: for detecting a still area in the video;

Vehicle identification unit: used to identify the stationary area by using the improved Haar classifier to achieve the purpose of detecting the vehicle; by training a plurality of strong classifiers, cascading the trained strong classifiers to form the improved Haar classifier, The input images are input one by one for classification and recognition, and after being recognized, they are input to the next-level classifier, and if they are not passed, they are directly recognized as non-vehicles.
The parking detection apparatus according to claim 9, wherein the apparatus further comprises: a shadow detecting unit for reducing the influence of the shadow on the foreground moving target extraction by using the HSV space-based shadow detecting method; the suspected still area detecting unit : used to establish a still pixel matrix. When a point is a foreground pixel, the value of the relative position of the suspected still pixel matrix is increased by one; if a point has a foreground into the background, the point is in the still pixel matrix. Is 0; if a point reaches the threshold set as expected, the point is considered to be a suspected still pixel, and the connected area consisting of suspected still pixels is a suspected still area.