Intelligent detection system and method for abnormal road spillage behavior
By combining image acquisition, recognition, and response modules and using multiple algorithms to preprocess images, the problem of detecting various types of projectiles in existing technologies has been solved, achieving high-precision projectile detection.
Patent Information
- Application Number
- CN202111602648.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-24
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2041-12-24
AI Technical Summary
Existing vision-based projectile detection systems can only detect a few types of projectiles, and training data is difficult to obtain, resulting in low robustness and difficulty in effectively detecting multiple types of projectiles.
The system, consisting of an image acquisition module, an image recognition module, and an information response module, preprocesses images using random erasure, Canny edge detection algorithm, feature extraction algorithm, and weather recognition algorithm, and uses optical flow method to determine abnormal spilling behavior.
It achieves high-precision detection of various types of projectiles, and can identify the size, spatial characteristics and motion state of the projectiles, thus improving the accuracy and robustness of the detection.
Smart Images

Figure CN114299002B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of abnormal road behavior detection, and in particular to an intelligent detection system and method for abnormal road spillage behavior. Background Art
[0002] In recent years, with the substantial increase in the number of cars, the occurrence of traffic accidents has also shown a clear upward trend. According to relevant statistics, driving with foreign objects thrown accounts for 22%-30% of traffic accidents and is also an important cause of fatal traffic accidents.
[0003] During high-speed driving, objects are often spilled due to human and non-human factors. To ensure safety on highways, currently, the mainly visual-based spilled object detection system is used to detect objects spilled on highways.
[0004] During the research and practice of the prior art, the inventors of the embodiments of the present application found that the existing vision-based spilled object detection systems are mostly target detection schemes. Since there are too many types of spilled objects and their sizes are uncertain, only a few types of spilled objects can be detected. In addition, training data is difficult to obtain and the robustness of the spilled object detection system is low. Summary of the invention
[0005] In order to overcome the shortcomings and deficiencies of the prior art, the present invention provides an intelligent detection system and method for abnormal road spillage behavior.
[0006] The technical solution adopted by the present invention is that the system is mainly composed of an image acquisition module, an image recognition module, and an information response module. First, a road monitoring camera is used to collect images of cars driving and the image information is input into a computer in real time for preprocessing. Then, the preprocessed images are recognized and detected using an algorithm.
[0007] The image acquisition module collects images through a monitoring camera in a road scene, performs key frame processing on the image, and then sends the key frame image data to the image recognition module; the function of the image recognition module is to detect abnormal behavior using an algorithm on the image; the function of the information response module is to perform mathematical statistics on the abnormal information and calculate the abnormal behavior vehicle information in the image.
[0008] Furthermore, a method for intelligently detecting abnormal behavior of road spillage is provided, and the steps are as follows:
[0009] Step S1: image preprocessing, using random erasing method to remove the occlusion of the target in the image;
[0010] Step S2: Image region segmentation: after image preprocessing, the image is denoised using a median filter algorithm to eliminate interference from virtual and real lines on the road surface, and then the Canny edge detection algorithm is used to segment different moving vehicles in the image;
[0011] Step S3: Extracting features of a small car, using a local binary pattern algorithm to extract texture features of the small car;
[0012] Step S4: extracting features of large trucks, using SIFT algorithm to extract texture features of large trucks;
[0013] Step S5: Weather feature recognition. Different weather conditions affect vehicle recognition effects. Weather conditions are recognized using contrast features and relative saturation.
[0014] Step S6: image comparison, using the optical flow method to identify the displacement position of the object in two adjacent images, and determine whether abnormal behavior of throwing occurs.
[0015] Furthermore, the random erasure method in step S1 has the following process:
[0016] Step S11: Set the erasure probability of the image to x. When the probability is greater than x, use a random area y in the image to fill the specified pixel value.
[0017] Step S12: Setting the parameters of the random area y, the area of the random area is represented by A, the value of A is the product of the number of vertical pixels and the number of horizontal pixels of the image, the height and width of the rectangle of the random erasing area are obtained by using the area of the random erasing area A1, and the calculation formula is:
[0018] A1=A*random(A2,A3),
[0019]
[0020]
[0021] Wherein, A2 and A3 represent the set erasure area thresholds, B1 represents the height of the erasure area, C1 represents the width of the erasure area, and b1 represents the aspect ratio of the erasure area;
[0022] Step S13: Randomly select a point Q1(m1, n1) in the image. The constraints of (m1, n1) are as follows:
[0023]
[0024] Where D m , D n , D1 and D3 represent the horizontal and vertical coordinate values and the length and width values of the upper left vertex of the image border respectively;
[0025] Step S14: assigning values to the randomly erased area and outputting the processed image.
[0026] Furthermore, the step S2 comprises the following specific steps:
[0027] S21: grayscale the image output in step S14;
[0028] S22: Perform median filtering denoising on the grayscale image, using a window size of 5 and pixels ×5;
[0029] S23: Using the Canny edge detection algorithm to extract image edge information, setting minimum and maximum thresholds, where the minimum threshold is 3 and the maximum threshold is 35;
[0030] S24: A binary image is obtained by using a dilation method for the image processed by the Canny edge detection algorithm. In each image, a point is found that is a connected area from a point to the bottom of the image, and the area above this point is set as a small car image, and the area below this point is set as a large truck image.
[0031] Furthermore, in step S3, the small car feature extraction is to divide the small car image into a plurality of small areas of size 2 and pixel × 2, and calculate the local binary value of each small area, the formula is:
[0032]
[0033] Where: (i f , j f ) represents the coordinates of the center pixel, 2h is the number of pixels, l f is the gray value of the central pixel, g represents the gth pixel adjacent to the central pixel, l h is the gray value of pixel g, p(i) is the sign function, and the formula is:
[0034]
[0035] Where i represents the independent variable of the symbolic function;
[0036] The local binary values of each region are counted, the histogram of each region is calculated and normalized, and the statistical histograms of each region are connected to form a complete feature vector, that is, the local binary features of the small car are obtained.
[0037] Furthermore, in step S4, the large truck feature extraction, the SIFT algorithm comprises the following three steps:
[0038] S41: Detection of extreme values in scale space, using the image to convolve with the Gaussian kernel function to obtain the Gaussian scale space S(i, j, ζ),
[0039] S(i,j,ζ)=T(i,j,ζ)V(i,j),
[0040] Where: i, j represent the horizontal and vertical coordinates of the pixel points in the image, ζ is the scale space factor, and different scale images are obtained by changing ζ. T(i, j, ζ) is the Gaussian kernel function, and V(i, j) represents the grayscale image. Its expression is:
[0041]
[0042] Construct the Gaussian difference scale space W(i, j, ζ),
[0043] W(i,j,ζ)=S(i,j,xζ)-S(i,j,ζ),
[0044] Where x represents the scale coefficient of the adjacent scale space factor. Compare the pixel points in the Gaussian difference scale space to determine whether they are extreme points. The local extreme points obtained are the feature points.
[0045] S42: Feature point positioning, using the fitting of a three-dimensional quadratic function to determine the scale and position of the feature point, while deleting the feature points with low contrast and unstable edge response points;
[0046] S43: Determine the direction of the feature point, and use the gradient direction of the pixels in the neighborhood of the feature point to obtain the direction value of the local structure.
[0047] Furthermore, in step S5, weather feature recognition is performed by extracting a ground area shadow set of sunny outdoor images from the image data set using a shadow detection algorithm, and calculating the distance between a single ground shadow and the shadow set shadow using a shadow distance calculation method, and extracting the most obvious ground shadow p of each sunny image in the data set. x , extract the most obvious ground shadow o of the image to be tested in the test set y , calculate the distance Z between a ground shadow in the image to be tested and the shadow in the shadow set p yx , the formula is:
[0048]
[0049] Furthermore, the weather contrast feature in step S5 uses the root mean square contrast as the global image feature, and the root mean square contrast Q m The calculation formula is:
[0050]
[0051] Where: Z (i,j) is the brightness of pixel (i, j); M t is the number of pixels in image t;
[0052] The calculation formula of the relative saturation is:
[0053]
[0054] Where: P i,j is the saturation of pixel t(i, j); max(P t ) is the maximum saturation of image t; min(P t ) is the minimum saturation of image t.
[0055] Furthermore, the optical flow method in step S6 determines the abnormal behavior of spilling by calculating the displacement of each pixel point of two adjacent images after a period of time △t. The constraint equation of the image is shown as follows:
[0056] A(i,j,t)=A(i+Δi,j+Δj,t+Δt),
[0057] Among them, A(i,j,t) represents the pixel at position (i,j) at time t.
[0058] Compared with the prior art, the present application proposes an intelligent detection system and method for abnormal road spillage behavior, which uses the system to capture images of a driving car, and pre-processes and processes the images through a series of algorithm flows such as random erasure method, Canny edge detection algorithm, feature extraction algorithm, weather recognition algorithm, and image comparison algorithm to determine whether abnormal spillage behavior occurs during the driving of the car. Compared with the prior art, the embodiment of the present application can analyze the actual foreign body state of the road surface with high precision and detect various types of spilled object categories, not limited to the size, spatial characteristics, category and motion state of the spilled objects. BRIEF DESCRIPTION OF THE DRAWINGS
[0059] Figure 1 is a system block diagram of the present invention;
[0060] Figure 2 The present invention is a flow chart of the method. DETAILED DESCRIPTION
[0061] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application may be combined with each other. The present application is further described in detail below in conjunction with the accompanying drawings and specific embodiments.
[0062] like Figure 1 As shown, a road spill abnormal behavior intelligent detection system and method, the system is mainly composed of an image acquisition module, an image recognition module, and an information response module. First, a road monitoring camera is used to collect images of cars driving and the image information is input into a computer in real time for preprocessing, and then the preprocessed images are recognized and detected using an algorithm;
[0063] The image acquisition module collects images through a monitoring camera in a road scene, performs key frame processing on the image, and then sends the key frame image data to the image recognition module; the function of the image recognition module is to detect abnormal behavior using an algorithm on the image; the function of the information response module is to perform mathematical statistics on the abnormal information and calculate the abnormal behavior vehicle information in the image.
[0064] The present invention adopts Hikvision DS-2DC4223IW-D / GLT(D) model camera, which supports a maximum image output of 1920×1080@30fps, uses H.265 compression algorithm, can greatly save storage space, and supports 5G network transmission.
[0065] The camera comes standard with an integrated braided cable, uses Ethernet communication, and uses an RJ45 network cable to connect hardware devices and switches. Compared with other communication methods, network cable communication is not affected by other communication waves, has strong anti-interference and high stability. Non-connected devices cannot obtain video information, and have a certain degree of confidentiality. In addition, a 100M network cable can meet the real-time nature of video transmission.
[0066] like Figure 2 As shown, a method for intelligently detecting abnormal road spilling behavior is shown, and the steps are as follows:
[0067] Step S1: image preprocessing, using random erasing method to remove the occlusion of the target in the image;
[0068] Step S2: Image region segmentation: after image preprocessing, the image is denoised using a median filter algorithm to eliminate interference from virtual and real lines on the road surface, and then the Canny edge detection algorithm is used to segment different moving vehicles in the image;
[0069] Step S3: Extracting features of a small car, using a local binary pattern algorithm to extract texture features of the small car;
[0070] Step S4: extracting features of large trucks, using SIFT algorithm to extract texture features of large trucks;
[0071] Step S5: Weather feature recognition. Different weather conditions affect vehicle recognition effects. Weather conditions are recognized using contrast features and relative saturation.
[0072] Step S6: image comparison, using the optical flow method to identify the displacement position of the object in two adjacent images, and determine whether abnormal behavior of throwing occurs.
[0073] Step S1 random erasure method, data enhancement is one of the commonly used methods for image preprocessing, which can expand samples, prevent overfitting, and improve model robustness to a certain extent. The present invention uses the random erasure data enhancement method to remove the situation where the target in the image is occluded, increase the number of data sets, and improve the generalization ability of the network, so that the network can complete the detection and recognition of the target only by learning local features. The process of this method is as follows:
[0074] Step S11: Set the erasure probability of the image to x. When the probability is greater than x, use a random area y in the image to fill the specified pixel value.
[0075] Step S12: Setting the parameters of the random area y, the area of the random area is represented by A, the value of A is the product of the number of vertical pixels and the number of horizontal pixels of the image, the height and width of the rectangle of the random erasing area are obtained by using the area of the random erasing area A1, and the calculation formula is:
[0076] A1=A*random(A2,A3),
[0077]
[0078]
[0079] Wherein, A2 and A3 represent the set erasure area thresholds, B1 represents the height of the erasure area, C1 represents the width of the erasure area, and b1 represents the aspect ratio of the erasure area;
[0080] Step S13: Randomly select a point Q1(m1, n1) in the image. The constraints of (m1, n1) are as follows:
[0081]
[0082] Where D m , D n , D1 and D3 represent the horizontal and vertical coordinate values and the length and width values of the upper left vertex of the image border respectively;
[0083] Step S14: assign a value to the erased area and output the processed image.
[0084] Car driving images can be divided into small cars and large trucks. Classifying the car types in the image and identifying small cars and large trucks are necessary tasks before feature extraction.
[0085] The present invention performs denoising on the image to eliminate the interference of virtual and real lines on the road surface. Then, based on the characteristics of high signal-to-noise ratio and detection accuracy of the edge detection of the Canny operator, the edge detection of the Canny operator is combined with morphology as the main method of skyline segmentation. The specific steps are:
[0086] S21: grayscale the image output in step S14;
[0087] S22: Perform median filtering denoising on the grayscale image using a window size of 5 and pixels ×5;
[0088] S23: Using the Canny edge detection algorithm to extract image edge information, setting minimum and maximum thresholds, where the minimum threshold is 3 and the maximum threshold is 35;
[0089] S24: A binary image is obtained by using a dilation method for the image processed by the Canny edge detection algorithm. In each image, a point is found that is a connected area from a point to the bottom of the image, and the area above this point is set as a small car image, and the area below this point is set as a large truck image.
[0090] Step S3: small car feature extraction: divide the small car image into multiple small areas of size 2 and pixel × 2, and calculate the local binary value of each small area. The formula is:
[0091]
[0092] Where: (i f , j f ) represents the coordinates of the center pixel, 2 h is the number of pixels, l f is the gray value of the central pixel, g represents the gth pixel adjacent to the central pixel, l h is the gray value of pixel g, p(i) is the sign function, and the formula is:
[0093]
[0094] Where i represents the independent variable of the symbolic function;
[0095] The local binary values of each region are counted, the histogram of each region is calculated and normalized, and the statistical histograms of each region are connected to form a complete feature vector, that is, the local binary features of the small car are obtained.
[0096] The feature points extracted by the SIFT algorithm remain unchanged to the influence of factors such as rotation, scale scaling, brightness change, etc., and have certain stability to perspective change, affine transformation, and noise. Based on the above factors, the SIFT algorithm is selected to extract the feature points of the large truck in the image. Step S4 is the large truck feature extraction. The SIFT algorithm includes the following 3 steps:
[0097] S41: Detection of extreme values in scale space, using the image to convolve with the Gaussian kernel function to obtain the Gaussian scale space S(i, j, ζ),
[0098] S(i,j,ζ)=T(i,j,ζ)V(i,j),
[0099] Among them, i, j represent the horizontal and vertical coordinates of the pixel points in the image, ζ is the scale space factor, and different scale images are obtained by changing ζ. T(i, j, ζ) is the Gaussian kernel function, and V(i, j) represents the grayscale image, which is expressed as:
[0100]
[0101] Construct the Gaussian difference scale space W(i, j, ζ),
[0102] W(i,j,ζ)=S(i,j,xζ)-S(i,j,ζ),
[0103] Where x represents the scale coefficient of the adjacent scale space factor. Compare the pixel points in the Gaussian difference scale space to determine whether they are extreme points. The local extreme points obtained are the feature points.
[0104] S42: Feature point positioning, using the fitting of a three-dimensional quadratic function to determine the scale and position of the feature point, while deleting the feature points with low contrast and unstable edge response points;
[0105] S43: Determine the direction of the feature point, and use the gradient direction of the pixels in the neighborhood of the feature point to obtain the direction value of the local structure.
[0106] Feature point description. First, correct the main rotation direction to ensure rotation invariance; then, generate a descriptor to form a 128-dimensional feature vector; finally, normalize the feature vector length.
[0107] Step S5: Weather feature recognition: Use shadow detection algorithm to extract ground area shadow set of sunny outdoor images from image data set, use shadow distance calculation method to calculate the distance between single ground shadow and shadow set shadow, extract the most obvious ground shadow p of each sunny image in data set x , extract the most obvious ground shadow o of the image to be tested in the test set y , calculate the distance Z between a ground shadow in the image to be tested and the shadow in the shadow set p yx , the formula is:
[0108]
[0109] Step S5: Weather contrast feature, using the root mean square contrast as the global image feature, the root mean square contrast Q m The calculation formula is:
[0110]
[0111] Where: Z (i,j) is the brightness of pixel (i, j); M t is the number of pixels in image t;
[0112] The calculation formula of the relative saturation is:
[0113]
[0114] Where: P i,j is the saturation of pixel t(i, j); max(P t ) is the maximum saturation of image t; min(P t ) is the minimum saturation of image t.
[0115] After extracting image features, random forest classifier is used for classification training. First, in order to make the training set of each tree different and contain repeated training samples, 10 training sets are extracted from the data set using Bootstrap sampling, each of which contains 10,000 car images; then k features (0 < k < 10) are randomly selected from the five features of SITF, local binary, shadow feature, contrast, and relative saturation, and classification nodes are selected in turn according to the information gain of each feature to establish a classification regression tree (CART). Repeat the above steps m times to establish m CART trees, and finally form a random forest with these m CART trees, establish a voting mechanism, and determine the category of each outdoor image.
[0116] Step S6: Optical flow method, by calculating the displacement of each pixel point of two adjacent images after time △t, the abnormal behavior of throwing is judged. The constraint equation of the image is shown as follows:
[0117] A(i,j,t)=A(i+Δi,j+Δj,t+Δt),
[0118] Among them, A(i,j,t) represents the pixel at position (i,j) at time t.
[0119] Compared with the prior art, the present application proposes an intelligent detection system and method for abnormal road spillage behavior, which uses the system to capture images of a driving car, and pre-processes and processes the images through a series of algorithm flows such as random erasure method, Canny edge detection algorithm, feature extraction algorithm, weather recognition algorithm, and image comparison algorithm to determine whether abnormal spillage behavior occurs during the driving of the car. Compared with the prior art, the embodiment of the present application can analyze the actual foreign body state of the road surface with high precision and detect various types of spilled object categories, not limited to the size, spatial characteristics, category and motion state of the spilled objects.
[0120] Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. An intelligent detection system for abnormal road spillage, characterized in that ,The system consists of image acquisition module, image recognition module and ,information response module. Firstly, the road monitoring camera is used to collect ,the images of the car driving and the image information is input into ,the computer in real time for preprocessing. Then the preprocessed images are ,recognized and detected using the algorithm; The image acquisition module collects images through a monitoring camera in a road scene, performs key frame processing on the images, and then sends the key frame image data to the image recognition module; The function of the image recognition module is to detect abnormal behavior using algorithms on images; The function of the information response module is to perform mathematical statistics on abnormal information and calculate the information of abnormal vehicles in the image; The system is implemented by an intelligent detection method for abnormal road spillage behavior, the steps of which are as follows: Step S1: image preprocessing, using random erasing method to remove the occlusion of the target in the image; Step S2: Image region segmentation: after image preprocessing, the image is denoised using a median filter algorithm to eliminate interference from virtual and real lines on the road surface, and then the Canny edge detection algorithm is used to segment different moving vehicles in the image; Step S3: Extracting features of a small car, using a local binary pattern algorithm to extract texture features of the small car; Step S4: extracting features of large trucks, using the SIFT algorithm to extract texture features of large trucks; Step S5: Weather feature recognition. Different weather conditions affect vehicle recognition effects. Weather conditions are recognized using contrast features and relative saturation. Step S6: image comparison, using the optical flow method to identify the displacement position of the object in two adjacent images, and determine whether there is abnormal behavior of throwing; The random erasure method in step S1 is as follows: Step S11: Set the erasure probability of the image to x. When the probability is greater than x, use a random area y in the image to fill the specified pixel value. Step S12: Setting the parameters of the random area y, the area of the random area is represented by A, the value of A is the product of the number of vertical pixels and the number of horizontal pixels of the image, the height and width of the rectangle of the random erasing area are obtained by using the area of the random erasing area A1, and the calculation formula is: A1=A*random(A2,A3), Wherein, A2 and A3 represent the set erasure area thresholds, B1 represents the height of the erasure area, C1 represents the width of the erasure area, and b1 represents the aspect ratio of the erasure area; Step S13: Randomly select a point Q1(m1, n1) in the image. The constraints of (m1, n1) are as follows: Where D m , D n , D1 and D3 represent the horizontal and vertical coordinate values and the length and width values of the upper left vertex of the image border respectively; Step S14: assigning values to the erased area and outputting the processed image; The step S2 comprises the following specific steps: S21: grayscale the image output in step S14; S22: Perform median filtering and denoising on the grayscale image; S23: Using the Canny edge detection algorithm to extract image edge information, setting minimum and maximum thresholds, where the minimum threshold is 3 and the maximum threshold is 35; S24: using the dilation method to obtain a binary image from the image processed by the Canny edge detection algorithm, in each image, find points that are all connected from a point to the bottom of the image, and set the area above this point to be a small car image, and the area below this point to be a large truck image; In step S3, the small car feature extraction is to divide the small car image into multiple small areas and calculate the local binary value of each small area. The formula is: Where: (i f , j f ) represents the coordinates of the center pixel, 2 h is the number of all pixels in the small area, l f is the gray value of the central pixel, g represents the gth pixel adjacent to the central pixel, l h is the gray value of pixel g, k is the sign function, and its value is 1 or 0; Count the local binary values of each region, calculate and normalize the histogram of each region, and connect the statistical histograms of each region to form a complete feature vector, that is, to obtain the local binary features of the small car; The step S4 is to extract the features of large trucks. The SIFT algorithm includes the following three steps: S41: Detection of extreme values in scale space, using the image to convolve with the Gaussian kernel function to obtain the Gaussian scale space S(i, j, ζ), S(i,j,ζ)=T(i,j,ζ)V(i,j), Where: i, j represent the horizontal and vertical coordinates of the pixel points in the image, ζ is the scale space factor, and different scale images are obtained by changing ζ. T(i, j, ζ) is the Gaussian kernel function, and V(i, j) represents the grayscale image. Its expression is: Construct the Gaussian difference scale space W(i, j, ζ), W(i,j,ζ)=S(i,j,xζ)-S(i,j,ζ), Where x represents the scale coefficient of the adjacent scale space factor. Compare the pixel points in the Gaussian difference scale space to determine whether they are extreme points. The local extreme points obtained are the feature points. S42: Feature point positioning, using the fitting of a three-dimensional quadratic function to determine the scale and position of the feature point, while deleting the feature points with low contrast and unstable edge response points; S43: Determine the direction of the feature point, and use the gradient direction of the pixels in the neighborhood of the feature point to obtain the direction value of the local structure; In step S5, weather feature recognition is performed by extracting a ground area shadow set of sunny outdoor images from the image data set using a shadow detection algorithm, calculating the distance between a single ground shadow and the shadow set shadow using a shadow distance calculation method, and extracting the most obvious ground shadow p of each sunny image in the data set. x , extract the most obvious ground shadow o of the image to be tested in the test set y , calculate the distance Z between a ground shadow in the image to be tested and the shadow in the shadow set p yx , the formula is: The weather contrast feature in step S5 is to use the root mean square contrast as the global image feature. m The calculation formula is: Where: Z (i,j) is the brightness of pixel (i, j); M t is the number of pixels in image t; The calculation formula of the relative saturation is: Where: P i,j is the saturation of pixel t(i, j); max(P t ) is the maximum saturation of image t; min(P t ) is the minimum saturation of image t; In the step S6 optical flow method, after extracting the image features, a random forest classifier is used for classification training, and k features are randomly selected from the five features of SIFT, local binary, shadow features, contrast, and relative saturation, and classification nodes are selected in turn according to the information gain of each feature to establish a classification regression tree; the above steps are repeated m times to establish m classification regression trees, and these m classification regression trees form a random forest, and a voting mechanism is established to determine the category of the scattered objects in each outdoor image, and the displacement of each pixel point of two adjacent images after a time △t is calculated. By calculating the displacement of each pixel point of two adjacent images after a time △t, the abnormal behavior of the spilling is judged, and the constraint equation of the image is shown as follows: A(i,j,t)=A(i+△i,j+△j,t+△t), Among them, A(i,j,t) represents the pixel at position (i,j) at time t.
Citation Information
Patent Citations
Road condition sensing abnormal event detection and alarm system of omni-directional detection radar and method thereof
CN109212520A
A pedestrian re-identification method based on a random erasure pedestrian alignment network
CN109583379A
Video-based fire smoke and thrown object detection method and system
CN112597975A
Method and device for detecting thrown objects on road surface and storage medium
CN113743151A
Abnormal behavior detecting apparatus and method thereof, and video monitoring system
US20120314064A1
