Real-time weak and small moving target detection filtering method based on rod cell bionics

Abstract

The embodiment of the invention provides a real-time weak moving target detection filtering method and device based on rod cell bionics. The method comprises the following steps: processing original to-be-processed image data based on a rod cell unit to obtain a G array; processing the array G by using a first moving target detection algorithm to obtain a first matrix, and processing the array G by using a second moving target detection algorithm to obtain a second matrix; processing the first matrix by using the improved apparent level cell unit to obtain a plurality of target points and target candidate areas, processing the plurality of target candidate areas by using the apparent bipolar cell unit to obtain a first area and a second area, and selecting the second area or the first area to obtain a final target candidate area; the weak and small moving target detection result is determined based on the second matrix and the final target candidate area, the filtering method helps to help the detection method to more stably detect the weak and small moving target under the complex dynamic background, and the false alarm rate and the missing report rate can be greatly reduced while the detection accuracy is guaranteed.

Description

Technical Field

[0001] The present application belongs to the field of motion target detection and signal processing, and in particular to a real-time weak motion target detection filtering method and device based on rod cell bionics. Background Art

[0002] For security systems, moving targets are more potentially dangerous. Some of these targets may be intrusive targets that move from non-sensitive areas outside the field of view to sensitive areas within the field of view, some may be suspicious targets that were originally stationary and lurking in the field of view and suddenly began to move, and some may be sensitive targets that need to be continuously tracked and monitored. Moving target detection has great application prospects in security and other fields, especially when the size of the moving target is very small and far away from the monitoring system. If the target can be detected and extracted in a timely and effective manner through the corresponding algorithm, it will provide sufficient and necessary preparation time for subsequent analysis and response and disposal work, thereby helping people to eliminate security threats and resolve potential crises in a timely and effective manner.

[0003] Prashant W.Patil et al. proposed an end-to-end moving target detection network. Compared with the comparison method, this method performs best in detection indicators such as prediction rate, recall rate, and F1 index. However, due to the characteristics of the deep convolutional neural network, this method is difficult to effectively detect weak moving targets. Li Songlin et al. proposed an improved small moving target detection network based on YoloV5. Compared with the comparison method, this method has better detection indicators for weak targets, but when the background is complex or changeable, the performance of the method is not stable enough. And due to the characteristics of YoloV5 itself, when the target is extremely small, this method still has the problem of target detection difficulty. Sajid Javed et al. invented a moving target detection model based on spatiotemporal sparse structure robust principal component analysis. Although this method can realize pixel-level moving target detection from a basic theory, and when the target is large, this method can effectively filter out background noise, but when the target is small, this method is likely to filter out the target by mistake. Summary of the invention

[0004] In view of this, the embodiments of the present application propose a real-time small moving target detection filtering method and device based on rod cell bionics, aiming to detect small moving targets more stably in complex dynamic backgrounds.

[0005] To achieve the above-mentioned purpose, an embodiment of the present application provides a real-time weak moving target detection filtering method based on rod cell bionics, comprising: obtaining original image data to be processed; obtaining a pre-constructed rod cell bionic model, wherein the rod cell bionic model includes a rod cell unit, an improved horizontal visual cell unit and a bipolar visual cell unit; processing the original image data to be processed based on the rod cell unit to obtain a high-dimensional spatiotemporal grayscale value information matrix; using a pre-constructed first moving target detection algorithm to process the high-dimensional spatiotemporal grayscale value information matrix to obtain a first matrix, and using a pre-constructed second moving target detection algorithm to process the high-dimensional spatiotemporal grayscale value information matrix to obtain a second matrix; using the improved horizontal visual cell unit to process the first matrix to obtain multiple target points and corresponding target candidate areas; using the bipolar visual cell unit to process each of the target candidate areas to obtain a first area and a second area, and selecting the second area or the first area based on preset conditions to obtain a final target candidate area; determining a weak moving target detection result based on the second matrix and the final target candidate area.

[0006] Optionally, the processing of the original image data to be processed based on the rod cell unit to obtain a high-dimensional spatiotemporal grayscale value information matrix includes: extracting the grayscale value of each image frame in the video data or the image sequence based on the rod cell unit; encoding the grayscale value of each image frame based on the rod cell unit to obtain the high-dimensional spatiotemporal grayscale value information matrix.

[0007] Optionally, the high-dimensional spatiotemporal grayscale value information matrix is ​​processed using a pre-constructed first motion target detection algorithm to obtain a first matrix, and the high-dimensional spatiotemporal grayscale value information matrix is ​​processed using a pre-constructed second motion target detection algorithm to obtain a second matrix, including: using an inexact augmented Lagrangian multiplier algorithm in which the weight of the sparse error term in the cost function is greater than 0.01 to process the high-dimensional spatiotemporal grayscale value information matrix to obtain the first matrix, wherein the first matrix is ​​used to characterize the foreground area with more significant motion, and the second matrix is ​​used to characterize the background area with aliasing noise and target details; using an inexact augmented Lagrangian multiplier algorithm in which the weight of the sparse error term in the cost function is less than 0.001 to process the high-dimensional spatiotemporal grayscale value information matrix to obtain the second matrix, wherein the second matrix is ​​used to characterize the foreground area containing boundaries, details and noise.

[0008] Optionally, each point in the first matrix is ​​an extreme foreground value point, the improved visual horizontal cell unit has a visual persistence unit and a combined clustering unit, and the method of using the improved visual horizontal cell unit to process the first matrix to obtain multiple target points and corresponding target candidate areas includes: obtaining multiple extreme foreground value points in the first matrix whose extreme foreground values ​​are greater than a preset value as target points; clustering each of the target points in the first matrix using the combined clustering unit to obtain corresponding first candidate areas; filtering all the extreme foreground value points in each of the first candidate areas according to a preset filtering rule to obtain each second candidate area; using the visual persistence unit to process the first matrix of multiple adjacent image frames adjacent to the current image frame, obtaining each neighborhood corresponding to each of the target points in each of the adjacent image frames, and determining multiple target candidate areas based on each of the target points and each of the neighborhoods in the current image frame.

[0009] Optionally, the value corresponding to each of the extreme foreground value points in the first matrix is ​​the extreme foreground value, and filtering all the extreme foreground value points in each of the first candidate regions according to a preset filtering rule to obtain each second candidate region includes: obtaining all the extreme foreground value points in each of the first candidate regions as target extreme foreground value points; obtaining multiple reference extreme foreground value points for each of the target extreme foreground value points, wherein multiple reference extreme foreground value points are determined based on points adjacent to the target extreme foreground value point in the horizontal and vertical directions; using a preset expression to process multiple reference extreme foreground value points to obtain filtering results of each of the current target extreme foreground value points; and determining each of the second candidate regions according to the filtering results of each of the target extreme foreground value points in each of the first candidate regions.

[0010] Optionally, the current extreme foreground value point in the first candidate area is (u, v), and the using a preset expression to process multiple reference extreme foreground value points to obtain the filtering results of each current target extreme foreground value point includes: using the preset expression to process multiple reference extreme foreground value points to obtain a zero value or a non-zero value calculation result of the expression, wherein the preset expression is:

[0011] Fe((u-1), v)||+||Fe((u+1), v)||)||(||Fe(u, (v-1))||+||Fe(u, (v+1))||)

[0012] Wherein, Fe((u-1), v) represents the extreme foreground value of the extreme foreground value point located on the left side of (u, v), Fe((u+1), v) represents the extreme foreground value of the extreme foreground value point located on the right side of (u, v), Fe(u, (v-1) represents the extreme foreground value of the extreme foreground value point located above (u, v), represents the extreme foreground value of the extreme foreground value point located on the right side of (u, v), + represents the AND operator, || represents the absolute value operator; if the calculation result of the expression is zero, the current extreme foreground value is determined to be zero; if the calculation result of the expression is non-zero, the current extreme foreground value is kept unchanged.

[0013] Optionally, based on each of the target points and each of the neighborhoods in the current image frame, a plurality of target candidate areas are determined, including: if the target point is within the range of each of the neighborhoods, then marking the second candidate area corresponding to the target point in the high-dimensional spatiotemporal grayscale value information matrix; if the target point exceeds the range of each of the neighborhoods, then abandoning the target point and marking the target point as a pre-selected target point, using the second candidate area corresponding to the pre-selected target point as data in the third candidate area, and using the visual persistence unit to search for a plurality of extreme foreground value points whose extreme foreground values ​​are greater than a preset value in the adjacent image frames as new target points, and using the combined clustering unit to process each of the new target points to obtain the corresponding fourth candidate areas; and using the K nearest neighbor clustering voting algorithm to perform cluster voting on the second candidate area, the third candidate area and the fourth candidate area to obtain each of the target candidate areas.

[0014] Optionally, the method of using a visual bipolar cell unit to process each of the target candidate regions to obtain a first region and a second region includes: obtaining grayscale information and texture information of each of the target candidate regions based on the visual bipolar cell unit; using a regional saliency clustering filtering algorithm to process the grayscale information and the texture information to obtain the first region and the second region of each of the target candidate regions, wherein the first region is a region where both the grayscale value and the brightness value are greater than a preset value, and the second region is a region where both the grayscale value and the texture value are less than a preset value.

[0015] Optionally, determining the small and weak moving target detection result based on the second matrix and the final target candidate area includes: performing an AND operation on the second matrix and the target candidate area to obtain a final small and weak moving target detection result.

[0016] To achieve the above-mentioned purpose, the embodiment of the present application also provides a real-time weak motion target detection filtering device based on rod cell bionics, comprising: a first acquisition module, used to acquire original image data to be processed; a second acquisition module, used to acquire a pre-constructed rod cell bionic model, wherein the rod cell bionic model includes a rod cell unit, an improved horizontal visual cell unit and a bipolar visual cell unit; a matrix construction module, used to process the original image data to be processed based on the rod cell unit to obtain a high-dimensional spatiotemporal grayscale value information matrix; a motion detection module, used to process the high-dimensional spatiotemporal grayscale value information using a pre-constructed first motion target detection algorithm A matrix is ​​obtained, and the high-dimensional spatiotemporal grayscale value information matrix is ​​processed by a pre-constructed second motion target detection algorithm to obtain a second matrix; a first processing module is used to process the first matrix by using an improved visual horizontal cell unit to obtain multiple target points and corresponding target candidate areas; a second processing module is used to process each of the target candidate areas by using a visual bipolar cell unit to obtain a first area and a second area, and select the second area or the first area based on a preset condition to obtain a final target candidate area; an output module is used to determine a weak motion target detection result based on the second matrix and the final target candidate area.

[0017] The embodiments of the present application propose a real-time weak small moving target detection filtering method and device based on rod cell bionics, which obtains original image data to be processed; obtains a pre-constructed rod cell bionic model, wherein the rod cell bionic model includes a rod cell unit, a horizontal cell unit and a bipolar cell unit; processes the original image data to be processed based on the rod cell unit to obtain a high-dimensional spatiotemporal grayscale value information matrix; uses a pre-constructed first moving target detection algorithm to process the high-dimensional spatiotemporal grayscale value information matrix to obtain a first matrix, and uses a pre-constructed second moving target detection algorithm to process the high-dimensional spatiotemporal grayscale value information matrix to obtain a second matrix; uses an improved horizontal cell unit to process the high-dimensional spatiotemporal grayscale value information matrix Processing the first matrix to obtain multiple target points and corresponding target candidate areas; using the visual bipolar cell unit to process the multiple target candidate areas to obtain the first area and the second area of ​​the target candidate area, and selecting the second area or the first area based on a preset condition to obtain the final target candidate area; determining the weak moving target detection result based on the second matrix and the final target candidate area. The filtering method of the present invention helps the detection method to detect weak moving targets more stably under complex dynamic backgrounds. Compared with a large number of comparison methods, the method of the present invention can significantly reduce the false alarm rate and the missed alarm rate while ensuring the detection accuracy, thereby improving the accuracy and stability level of the detection work. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1is a flow chart of a rod cell bionic real-time weak moving target detection filtering method provided in one embodiment of the present application;

[0019] Figure 2 This is a process of a rod cell bionic real-time weak moving target detection filtering method provided in an embodiment of the present application. Figure 2 ;

[0020] Figure 3 It is a visualization schematic diagram of an expression of a rod cell bionic real-time weak moving target detection filtering method provided in one embodiment of the present application;

[0021] Figure 4 It is each detection image in the detection process of the rod cell bionic real-time weak moving target detection filtering method provided in one embodiment of the present application;

[0022] Figure 5 This is a structural block diagram of a rod cell bionic real-time weak motion target detection filtering device provided by another embodiment of the present application. DETAILED DESCRIPTION

[0023] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the embodiments of the present application will be described in detail below in conjunction with the accompanying drawings. However, it will be appreciated by those skilled in the art that in the present application, many technical details are proposed in order to enable the reader to better understand the present application. However, even without these technical details and various changes and modifications based on the following embodiments, the technical scheme claimed in the present application can also be implemented. The division of the following embodiments is for the convenience of description, and the specific implementation of the present application should not be construed as any limitation, and the various embodiments can be combined and referenced with each other under the premise of no contradiction.

[0024] Surveillance cameras generally have many background elements, many background changes, and poor field of view stability during operation, which cause a lot of and serious interference to the detection work. Unlike moving targets of ordinary size, small targets have too few pixels, too little information, and insufficient motion characteristics. Small targets are not only difficult to effectively distinguish from the background and other moving foreground areas, but also difficult to effectively characterize. These will seriously interfere with the normal operation of the detection method, reducing the accuracy, underreporting rate, and false alarm rate of the detection results. In other words, the existing moving target detection model is limited by the influence of weather, lighting, clouds, natural background changes, camera base movement, and input data noise, and it is very easy to mix a large amount of noise that is difficult to filter out into the detection results.

[0025] In order to solve the above-mentioned problems existing in the prior art, the present invention provides a real-time small moving target detection filtering method based on rod cell bionics. The method utilizes the rod cell bionics principle to design a real-time small moving target detection filtering method, which can quickly and efficiently filter out noise interference while effectively retaining the foreground information of small targets. At the same time, multiple indicator dimensions such as accuracy, prediction rate, recall rate, F1 index, etc. will be improved, thereby effectively solving the problem of difficulty in detecting small moving targets in complex backgrounds.

[0026] It should be noted that the real-time weak moving target detection filtering method based on rod cell bionics can be applied to electronic devices, wherein the electronic device can be a terminal or a server, and this embodiment and the following embodiments are all described by taking the server as an example. The implementation details of the real-time weak moving target detection filtering method based on rod cell bionics proposed in this embodiment are specifically described below. The following content is only for the convenience of understanding the implementation details provided, and is not necessary for the implementation of this solution.

[0027] The specific process of the rod cell bionic real-time weak moving target detection filtering method proposed in this embodiment can be as follows: Figure 1 As shown, the real-time weak moving target detection filtering method based on rod cell bionics includes:

[0028] S101, obtaining original image data to be processed;

[0029] Specifically, the server reads the input video or time-sequential image frame sequence data, and the input module aligns the data in time and space and then passes it to the rod cell unit for processing.

[0030] S102, obtaining a pre-constructed bionic model of a rod cell, wherein the bionic model of a rod cell includes a rod cell unit, an improved horizontal cell unit, and a bipolar cell unit;

[0031] S103, processing the original image data to be processed based on the rod cell unit to obtain a high-dimensional spatiotemporal grayscale value information matrix;

[0032] In an embodiment of the present application, step S103 may include the following execution process:

[0033] S1031, extracting the grayscale value of each image frame in the video data or the image sequence based on the rod cell unit;

[0034] S1032. Encode the grayscale value of each image frame based on the rod cell unit to obtain the high-dimensional spatiotemporal grayscale value information matrix.

[0035] refer to Figure 2 and Figure 1In the specific implementation process, the server uses the rod cell unit to process the received spatiotemporally aligned video or time-sequenced image frame sequence data to obtain its grayscale value information, and encodes the grayscale value information into a high-dimensional spatiotemporal grayscale value information matrix (G matrix) according to the time sequence and spatial relationship of the data. The cross-section of the G matrix along the time dimension is the spatial grayscale value information of the original video or image frame sequence data at that moment.

[0036] S104, using a pre-constructed first moving target detection algorithm to process the high-dimensional spatiotemporal grayscale value information matrix to obtain a first matrix, and using a pre-constructed second moving target detection algorithm to process the high-dimensional spatiotemporal grayscale value information matrix to obtain a second matrix;

[0037] In an embodiment of the present application, step S104 may include the following execution process:

[0038] S1041, using an inexact augmented Lagrange multiplier algorithm in which the weight of the sparse error term in the cost function is greater than 0.01 to process the high-dimensional spatiotemporal grayscale value information matrix to obtain a first matrix, wherein the first matrix is ​​used to characterize a foreground area with more significant motion;

[0039] S1042. Use an inexact augmented Lagrange multiplier algorithm in which the weight of the sparse error term in the cost function is less than 0.001 to process the high-dimensional spatiotemporal grayscale value information matrix to obtain a second matrix, wherein the second matrix is ​​used to characterize the foreground area including boundaries, details and noise.

[0040] refer to Figure 2 and Figure 1 In the specific implementation process, the rod cell unit will hand over the obtained high-dimensional spatiotemporal grayscale value information matrix to two target detection algorithms for processing.

[0041] Specifically, the server can divide the high-dimensional spatiotemporal grayscale value information matrix by executing steps S1041 to S1042. The motion target detection unit can be composed of two independent motion target detection algorithms that run simultaneously, and the two target detection algorithms are respectively referred to as the first motion target detection algorithm and the second motion target detection algorithm. Among them, the first motion target detection algorithm adopts a coarse-grained motion target detection model, which is characterized in that the input data can be divided into a foreground area with more significant motion (Fe matrix, i.e., the first matrix, each point in the Fe matrix is ​​called an extreme foreground value point, and its corresponding Fe value is the extreme foreground value) and a background area with a large amount of aliasing noise and target details (Br matrix, each point in the Br matrix is ​​called a rough background value point, and its corresponding Br value is a rough background value). In this specific embodiment, the inexact augmented Lagrange multiplier method with a large weight (λ=0.01) of the sparse error term in the cost function is selected to divide the input G matrix into the Fe matrix representing the foreground and the Br matrix representing the background; the second operation The moving target detection algorithm uses a fine-grained moving target detection model to divide the input data into a foreground area (Fr matrix, i.e., the second matrix, each point in the Fr matrix is ​​called a rough foreground value point, and its corresponding Fr value is a rough foreground value) containing precise boundaries and rich details but aliased with a large amount of noise that is difficult to filter out, and a relatively stable background area (Be matrix, each point in the Be matrix is ​​called an extreme background value point, and its corresponding Be value is an extreme background value). This specific embodiment uses an inexact augmented Lagrange multiplier method with a small weight (λ=0.001) of the sparse error term in the cost function. This method can divide the input G matrix into an Fr matrix representing the foreground and a Be matrix representing the background. At the same time, since the running speed of the first moving target detection algorithm is significantly faster than that of the second moving target detection algorithm, during the running of the second moving target detection algorithm, the detection results generated by the first moving target detection algorithm can be further processed at the same time.

[0042] S105, using the improved visual horizontal cell unit to process the first matrix to obtain multiple target points and corresponding target candidate areas;

[0043] In this step, the server passes the first matrix obtained after processing by the target detection unit to the visual horizontal cell unit for processing. The server first obtains multiple target points and corresponding candidate areas in the first matrix through the visual horizontal cell unit, and then comprehensively analyzes the relevant information of multiple candidate areas and the G matrix related information under the action of the visual link, and finally gives the target candidate area to be determined.

[0044] Each point in the first matrix is ​​an extreme foreground value point. The improved visual horizontal cell unit has a visual persistence unit and a combined clustering unit. In the embodiment of the present application, step S105 may include the following execution process:

[0045] S1051, obtaining a plurality of extreme foreground value points in the first matrix whose extreme foreground values ​​are greater than a preset value as target points;

[0046] Specifically, the horizontal visual cell first finds several target points with extremely significant motion (several points with larger extreme foreground values) from the Fe array given by the motion target detection model 1 as the target points to be determined.

[0047] S1052, clustering each target point in the first matrix using a combined clustering unit to obtain corresponding first candidate regions;

[0048] Among them, the algorithm used by the combined clustering unit to cluster each target point in the current image frame is grayscale-based neighborhood continuity clustering. The combined clustering unit performs neighborhood continuity clustering on the Fe array near each target point, and combines the size of the clustering result and the prior information of the target (the corresponding information of the previous frames) to determine the specific size of the square neighborhood of the point as the first candidate area (the area size can be referenced as [5, 5], [10, 10], [15, 15], [20, 20], [25, 25], [30, 30], etc., and [50, 50] is used for this example sample). At the same time, in order to effectively filter out a large amount of widely existing scattered noise, a preset filtering rule is designed to further improve the positioning accuracy of the target positioning point and the accuracy of the target candidate area.

[0049] S1053, filtering all extreme foreground value points in each first candidate region according to a preset filtering rule to obtain each second candidate region;

[0050] Among them, the server sends each foreground value point and its neighborhood information in the candidate target area of ​​the Fe array into a preset expression for calculation through the visual horizontal cell unit, and filters out the relevant foreground value points whose calculation results do not meet the preset filtering rules to obtain the second candidate area. This can effectively filter out the abnormal scattered noise in the first candidate area, thereby reducing interference with the detection results.

[0051] Specifically, step S1053 may include the following execution process:

[0052] S10531, obtaining all extreme foreground value points in each first candidate area as target extreme foreground value points;

[0053] S10532, obtaining a plurality of reference extreme foreground value points for each target extreme foreground value point, wherein the plurality of reference extreme foreground value points are determined based on each point adjacent to the target extreme foreground value point in the horizontal axis direction and the vertical axis direction;

[0054] S10533, using a preset expression to process multiple reference extreme foreground value points to obtain filtering results of current target extreme foreground value points;

[0055] Wherein, the current extreme foreground value point in the first candidate region is (u, v), and step S10533 may include processing a plurality of reference extreme foreground value points using a preset expression to obtain a zero value or a non-zero value calculation result of the expression, wherein the preset expression is:

[0056] Fe((u-1), v)||+||Fe((u+1), v)||)||(||Fe(u, (v-1))||+||Fe(u, (v+1))||)

[0057] Among them, Fe((u-1), v) represents the extreme foreground value of the extreme foreground value point located on the left side of (u, v), Fe((u+1), v) represents the extreme foreground value of the extreme foreground value point located on the right side of (u, v), Fe(u, (v-1)) represents the extreme foreground value of the extreme foreground value point located above (u, v), represents the extreme foreground value of the extreme foreground value point located on the right side of (u, v), + represents the AND operator, || represents the absolute value operator; if the calculation result of the expression is zero, the current extreme foreground value is determined to be zero; if the calculation result of the expression is non-zero, the current extreme foreground value remains unchanged.

[0058] In the specific execution process, the server uses the four extreme foreground value points adjacent to the current extreme foreground point along the horizontal axis and the vertical axis as its neighborhood reference extreme foreground area, and other areas are its neighborhood non-reference extreme foreground areas. Suppose the current extreme foreground value point in the first candidate area is (u, v), and its extreme foreground value is Fe(u, v); the reference extreme foreground value point on its left is ((u-1), v), and its extreme foreground value is Fe((u-1), v); the reference extreme foreground value point on its right is ((u+1), v), and its extreme foreground value is Fe((u+1), v); the reference extreme foreground value point on its upper side is (u, (v-1)), and its extreme foreground value is Fe( u, (v-1)); the reference extreme foreground value point on its lower side is (u, (v+1)), and its extreme foreground value is Fe(u, (v+1)), then the filtering calculation expression is (||Fe((u-1), v)||+||Fe((u+1), v)||)||(||Fe(u, (v-1))||+||Fe(u, (v+1))||), that is, the absolute value of Fe((u-1), v) is ANDed with the absolute value of Fe((u+1), v), the absolute value of Fe(u, (v-1)) is ANDed with the absolute value of Fe(u, (v+1)), and then the results of the two AND operations are ORed, where the visualization of the expression is as follows Figure 3 As shown, from left to right, the first figure is the calculation in the horizontal axis direction, the second figure is the calculation in the vertical axis direction, and the third figure is the calculation result.

[0059] S10534. Determine each second candidate region according to the filtering result of each target extreme foreground value point in each first candidate region.

[0060] In other words, when and only when the result of the above expression is 0, the previous extreme foreground value of the current extreme foreground value point is set to Fe(u, v) = 0, that is, the point of each first candidate region is filtered out to obtain each second candidate region. The above operation can effectively filter out the scattered interference noise that may be mixed into the Fe array.

[0061] S1054, using the persistence of vision unit to process the first matrix of multiple adjacent image frames adjacent to the current image frame, obtain each neighborhood corresponding to each target point in each adjacent image frame, and determine multiple target candidate areas based on each target point and each neighborhood in the current image frame.

[0062] It is worth noting that the present application also designs a persistence of vision unit based on the persistence of vision effect. The principle is to construct a persistence of vision unit based on the persistence function of animal vision. The persistence of vision unit will memorize the position information of the target and its neighborhood information in the previous few frames.

[0063] In an embodiment of the present application, step S105 may include the following execution process:

[0064] Determine whether the target point is within the range of each neighborhood;

[0065] If the target point is within the range of each neighborhood, the second candidate area corresponding to the target point is marked in the high-dimensional spatiotemporal gray value information matrix;

[0066] If the target point exceeds the range of each neighborhood, the target point is abandoned and marked as a pre-selected target point, the second candidate area corresponding to the pre-selected target point is used as the data in the third candidate area, and the visual persistence unit is used to search for multiple extreme foreground value points with extreme foreground values ​​greater than a preset value in adjacent image frames as new target points, and the combined clustering unit is used to process each new target point to obtain each corresponding fourth candidate area;

[0067] The K nearest neighbor cluster voting algorithm is used to perform cluster voting on the second candidate area, the third candidate area and the fourth candidate area to obtain each target candidate area.

[0068] In the specific execution process, when the target point (i.e., the target point) given by the Fe array is within a certain neighborhood range of the target point in the previous frames, it is considered that the current target positioning is accurate, and the target point (X, Y) is marked in the current time section in the G array, and the neighborhood information of the target point (X, Y) in the G array is sent to the visual bipolar cell unit for further analysis and judgment; when the target point (X, Y) given by the Fe array far exceeds the neighborhood range of the target point in the previous frames, it is determined that the current target positioning information is incorrect, and the target point is determined to be the pre-selected target point (X-Pre, Y-Pre), and the extremely significant motion area is re-searched in the neighborhood of the target point in the previous frames, and the target point, i.e., the new target point, is re-determined, and the neighborhood information of the target point (X, Y), the pre-selected target point (X-Pre, Y-Pre) and the new target point in the G array and the Fe array is subjected to K nearest neighbor clustering voting to obtain the final target point, and the target candidate area of ​​the target point in the current time section in the G array and the Fe array is sent to the visual bipolar cell unit for further analysis and judgment.

[0069] S106, using the visual bipolar cell unit to process each target candidate region to obtain a first region and a second region, and selecting the second region or the first region based on a preset condition to obtain a final target candidate region;

[0070] S107: Determine a small moving target detection result based on the second matrix and the final target candidate region.

[0071] In an embodiment of the present application, step S107 may include the following execution process:

[0072] S1071, obtaining grayscale information and texture information of each target candidate area based on the visual bipolar cell unit;

[0073] S1072. Use the regional saliency clustering filtering algorithm to process the grayscale information and texture information to obtain the first area and the second area of ​​each target candidate area, wherein the first area is an area where both the grayscale value and the brightness value are greater than the preset value, and the second area is an area where both the grayscale value and the texture value are less than the preset value.

[0074] Specifically, the server uses the visual link of the bipolar cell unit to perform regional saliency clustering filtering on the incoming relevant information combined with its grayscale texture changes, and combines the grayscale texture information of the target to divide the clustering results into bright areas (On areas) and dark areas (Off areas), and then obtains the final target candidate area through regional selection.

[0075] After step S107, the server saves and visualizes the detection results.

[0076] refer to Figure 4 ,From right to left, the three images are the image before detection, the detected background image, and the detected foreground image.

[0077] In summary, the rod cell bionic real-time weak moving target detection filtering method provided by the above-mentioned embodiments of the present application is through obtaining the original image data to be processed; obtaining a pre-constructed rod cell bionic model, wherein the rod cell bionic model includes a rod cell unit, a horizontal visual cell unit and a bipolar visual cell unit; processing the original image data to be processed based on the rod cell unit to obtain a high-dimensional spatiotemporal grayscale value information matrix; using a pre-constructed first moving target detection algorithm to process the high-dimensional spatiotemporal grayscale value information matrix to obtain a first matrix, and using a pre-constructed second moving target detection algorithm to process the high-dimensional spatiotemporal grayscale value information matrix to obtain a second matrix; using an improved horizontal visual cell unit The first matrix is ​​processed by the element to obtain multiple target points and corresponding target candidate areas; the multiple target candidate areas are processed by the visual bipolar cell unit to obtain the first area and the second area of ​​the target candidate area, and the second area or the first area is selected based on a preset condition to obtain the final target candidate area; the weak moving target detection result is determined based on the second matrix and the final target candidate area. The filtering method of the present invention helps the detection method to detect weak moving targets more stably under complex dynamic backgrounds. Compared with a large number of comparison methods, the method of the present invention can significantly reduce the false alarm rate and the missed alarm rate while ensuring the detection accuracy, thereby improving the accuracy and stability level of the detection work.

[0078] It should be noted that the step division of the above methods is only for clear description. When implemented, they can be combined into one step or some steps can be split and decomposed into multiple steps. As long as they include the same logical relationship, they are all within the scope of protection of this application; adding insignificant modifications to the algorithm or process or introducing insignificant designs without changing the core design of the algorithm and process are all within the scope of protection of this application.

[0079] In the above embodiment, another embodiment of the present application proposes a real-time weak moving target detection filtering device based on rod cell bionics, referring to Figure 5, the detection filter device 100 may include a first acquisition module 1001, a second acquisition module 1002, a matrix construction module 1003, a motion detection module 1004, a first processing module 1005, a second processing module 1006 and an output module 1007: the first acquisition module 1001 is used to acquire the original image data to be processed; the second acquisition module 1002 is used to acquire a pre-constructed rod cell bionic model, wherein the rod cell bionic model includes a rod cell unit, an improved horizontal cell unit and a bipolar cell unit; the matrix construction module 1003 is used to process the original image data to be processed based on the rod cell unit to obtain a high-dimensional spatiotemporal grayscale value information matrix; the motion detection module 1004 is used to utilize A high-dimensional spatiotemporal grayscale value information matrix is ​​processed by a pre-constructed first motion target detection algorithm to obtain a first matrix, and a high-dimensional spatiotemporal grayscale value information matrix is ​​processed by a pre-constructed second motion target detection algorithm to obtain a second matrix; a first processing module 1005 is used to process the first matrix using an improved visual horizontal cell unit to obtain multiple target points and corresponding target candidate areas; a second processing module 1006 is used to process each target candidate area using a visual bipolar cell unit to obtain a first area and a second area, and select the second area or the first area based on preset conditions to obtain a final target candidate area; an output module 1007 is used to determine a weak motion target detection result based on the second matrix and the final target candidate area.

[0080] It is not difficult to find that this embodiment is a system embodiment corresponding to the above method embodiment, and this embodiment can be implemented in conjunction with the above method embodiment. The relevant technical details and technical effects mentioned in the above embodiments are still valid in this embodiment, and in order to reduce repetition, they are not repeated here. Accordingly, the relevant technical details mentioned in this embodiment can also be applied in the above embodiments.

[0081] It is worth mentioning that all modules involved in this embodiment are logic modules. In practical applications, a logic unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. In addition, in order to highlight the innovative part of this application, this embodiment does not introduce units that are not closely related to solving the technical problems proposed by this application, but this does not mean that there are no other units in this embodiment.

[0082] Those skilled in the art will appreciate that the above embodiments are specific embodiments for implementing the present application, and in actual applications, various changes may be made thereto in form and detail without departing from the spirit and scope of the present application.

Claims

1. A real-time weak moving target detection filtering method based on rod cell bionics, characterized in that: include: Obtaining original image data to be processed; Acquire a pre-built bionic model of a rod cell, wherein the bionic model of a rod cell comprises a rod cell unit, an improved horizontal cell unit and a bipolar cell unit; Processing the original image data to be processed based on the rod cell unit to obtain a high-dimensional spatiotemporal grayscale value information matrix; Processing the high-dimensional spatiotemporal grayscale value information matrix using a pre-constructed first moving target detection algorithm to obtain a first matrix, and processing the high-dimensional spatiotemporal grayscale value information matrix using a pre-constructed second moving target detection algorithm to obtain a second matrix; Processing the first matrix using the improved visual horizontal cell unit to obtain a plurality of target points and corresponding target candidate regions; Processing each of the target candidate regions using the visual bipolar cell unit to obtain a first region and a second region, and selecting the second region or the first region based on a preset condition to obtain a final target candidate region; A small moving target detection result is determined based on the second matrix and the final target candidate area.

2. The rod cell bionic real-time small moving target detection filtering method according to claim 1 is characterized in that: The step of processing the original image data to be processed based on the rod cell unit to obtain a high-dimensional spatiotemporal grayscale value information matrix includes: Extracting the grayscale value of each image frame in the video data or the image sequence based on the rod cell unit; The grayscale value of each image frame is encoded based on the rod cell unit to obtain the high-dimensional spatiotemporal grayscale value information matrix.

3. The rod cell bionic real-time small moving target detection filtering method according to claim 1 is characterized in that: The method of using a pre-built first moving target detection algorithm to process the high-dimensional spatiotemporal grayscale value information matrix to obtain a first matrix, and using a pre-built second moving target detection algorithm to process the high-dimensional spatiotemporal grayscale value information matrix to obtain a second matrix includes: Using an inexact augmented Lagrange multiplier algorithm with a weight of a sparse error term in a cost function greater than 0.01 to process the high-dimensional spatiotemporal grayscale value information matrix to obtain a first matrix, wherein the first matrix is ​​used to characterize a foreground area with more significant motion; The high-dimensional spatiotemporal grayscale value information matrix is ​​processed by an inexact augmented Lagrange multiplier algorithm in which the weight of the sparse error term in the cost function is less than 0.001 to obtain a second matrix, wherein the second matrix is ​​used to characterize a foreground area including boundaries, details and noise.

4. The rod cell bionic real-time small moving target detection filtering method according to claim 1, characterized in that: Each point in the first matrix is ​​an extreme foreground value point, the improved visual horizontal cell unit has a visual persistence unit and a combined clustering unit, and the first matrix is ​​processed by using the improved visual horizontal cell unit to obtain multiple target points and corresponding target candidate areas, including: Acquire a plurality of extreme foreground value points in the first matrix whose extreme foreground values ​​are greater than a preset value as target points; Clustering the target points in the first matrix using a combined clustering unit to obtain corresponding first candidate regions; Performing filtering processing on all extreme foreground value points in each of the first candidate regions according to a preset filtering rule to obtain each of the second candidate regions; The visual persistence unit is used to process a first matrix of multiple adjacent image frames adjacent to the current image frame, and each neighborhood corresponding to each target point is obtained in each of the adjacent image frames, and multiple target candidate areas are determined based on each of the target points and each of the neighborhoods in the current image frame.

5. The rod cell bionic real-time small moving target detection filtering method according to claim 4 is characterized in that: The value corresponding to each of the extreme foreground value points in the first matrix is ​​an extreme foreground value, and filtering all the extreme foreground value points in each of the first candidate regions according to a preset filtering rule to obtain each second candidate region includes: Acquire all extreme foreground value points in each of the first candidate regions as target extreme foreground value points; Acquire multiple reference extreme foreground value points of each of the target extreme foreground value points, wherein the multiple reference extreme foreground value points are determined based on points adjacent to the target extreme foreground value point in the horizontal and vertical directions; Processing a plurality of reference extreme foreground value points using a preset expression to obtain filtering results of each of the current target extreme foreground value points; The second candidate regions are determined according to the filtering results of the target extreme foreground value points in the first candidate regions.

6. The rod cell bionic real-time small moving target detection filtering method according to claim 5, characterized in that: The current extreme foreground value point in the first candidate area is (u, v), and the method of processing a plurality of reference extreme foreground value points using a preset expression to obtain filtering results of each of the current target extreme foreground value points includes: The preset expression is used to process the plurality of reference extreme foreground value points to obtain a zero value or a non-zero value calculation result of the expression, wherein the preset expression is: (||Fe((u-1), v)||+||Fe((u+1), v)||)||(||Fe(u, (v-1))||+||Fe(u, (v+1))||) Wherein, Fe((u-1), v) represents the extreme foreground value of the extreme foreground value point on the left side of (u, v), Fe((u+1), v) represents the extreme foreground value of the extreme foreground value point on the right side of (u, v), Fe(u, (v-1)) represents the extreme foreground value of the extreme foreground value point above (u, v), represents the extreme foreground value of the extreme foreground value point on the right side of (u, v), + represents the AND operator, and |||| represents the absolute value operator; If the calculation result of the expression is zero, then the current extreme prospect value is determined to be zero; If the calculation result of the expression is a non-zero value, the current extreme prospect value is kept unchanged.

7. The rod cell bionic real-time small moving target detection filtering method according to claim 5, characterized in that: The determining of a plurality of target candidate regions based on each of the target points and each of the neighborhoods in the current image frame comprises: Determine whether the target point is within the range of each of the neighborhoods; If the target point is within the range of each of the neighborhoods, marking the second candidate area corresponding to the target point in the high-dimensional spatiotemporal grayscale value information matrix; If the target point exceeds the range of each of the neighborhoods, the target point is abandoned and marked as a pre-selected target point, the second candidate area corresponding to the pre-selected target point is used as data in the third candidate area, and the visual persistence unit is used to search for multiple extreme foreground value points with extreme foreground values ​​greater than a preset value in the adjacent image frames as new target points, and the combined clustering unit is used to process each of the new target points to obtain corresponding fourth candidate areas; The second candidate region, the third candidate region and the fourth candidate region are clustered and voted on using a K-nearest neighbor cluster voting algorithm to obtain the target candidate regions.

8. The rod cell bionic real-time small moving target detection filtering method according to claim 1, characterized in that: The method of processing each of the target candidate regions by using the visual bipolar cell unit to obtain the first region and the second region includes: Acquire grayscale information and texture information of each of the target candidate regions based on the visual bipolar cell unit; The grayscale information and the texture information are processed using a regional saliency clustering filtering algorithm to obtain a first area and a second area of ​​each target candidate area, wherein the first area is an area where both the grayscale value and the brightness value are greater than a preset value, and the second area is an area where both the grayscale value and the texture value are less than a preset value.

9. The rod cell bionic real-time small moving target detection filtering method according to claim 1, characterized in that: The determining of the weak moving target detection result based on the second matrix and the final target candidate area includes: An AND operation is performed on the second matrix and the target candidate area to obtain a final weak moving target detection result.

10. A real-time weak moving target detection filter device based on rod cell bionics, characterized in that: include: A first acquisition module, used for acquiring original image data to be processed; A second acquisition module is used to acquire a pre-constructed bionic model of a rod cell, wherein the bionic model of a rod cell includes a rod cell unit, an improved horizontal cell unit and a bipolar cell unit; A matrix construction module, used for processing the original image data to be processed based on the rod cell unit to obtain a high-dimensional spatiotemporal grayscale value information matrix; A motion detection module, configured to process the high-dimensional spatiotemporal grayscale value information matrix using a pre-constructed first motion target detection algorithm to obtain a first matrix, and to process the high-dimensional spatiotemporal grayscale value information matrix using a pre-constructed second motion target detection algorithm to obtain a second matrix; A first processing module, configured to process the first matrix using the improved visual horizontal cell unit to obtain a plurality of target points and corresponding target candidate regions; A second processing module, configured to process each of the target candidate regions using the visual bipolar cell unit to obtain a first region and a second region, and select the second region or the first region based on a preset condition to obtain a final target candidate region; An output module is used to determine a weak moving target detection result based on the second matrix and the final target candidate area.

Images