Video moving target detection method
Through the improved robust non-convex norm model, the low-rank sum sparse term is approximate using the non-convex substitution function, and the noise term is introduced, combined with the alternating direction multiplier method, the problem of poor detection of video motion targets in complex scenarios is solved. High-precision and low-complexity video motion object detection is achieved.
Patent Information
- Application Number
- CN202411406453.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-10
- Publication Date
- 2024-12-20
AI Technical Summary
The existing video motion object detection method is not effective in complex scenarios, especially the fact that it has not effectively dealt with factors such as noise, lighting changes and camera jitter. It is difficult to directly solve the NP-hard problem of low-rank sparse decomposition model.
Using an improved robust non-convex norm video motion object detection model, replace the low-rank term and non-convex Geman fraction norm with non-convex, and introduce noise terms, combined with alternating The direction multiplication method is used to solve it.
The accuracy and noise resistance of video motion object detection are improved, and its superiority is verified from a visual and quantitative perspective, with good video motion object detection performance and low time complexity.
Smart Images
Figure CN119169511A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of multimedia processing, and in particular relates to a video moving target detection method. Background Art
[0002] The purpose of video moving target detection is to decompose the video into dynamic foreground targets and static background. It is the first step in the video surveillance system and is widely used in many fields such as social security, traffic management, and medical health. The performance of video moving target detection is directly related to the performance of downstream tasks, such as target tracking and target recognition. However, due to factors such as noise, lighting changes, and camera jitter in the video, video moving target detection faces huge challenges.
[0003] In the middle of the last century, most traditional video moving target detection models used pixel-level background modeling methods, such as Gaussian mixture model, Vibe model and principal component analysis (PCA). Although these methods can extract moving targets in some simple scenes, they completely ignore the correlation between video pixels, making it difficult to adapt to video moving target detection in complex scenes. To address the above problems, the Low Rank and Sparse Decomposition (LRSD) model was proposed and applied to video moving target detection. This model recognizes that the background sequence of the video changes slowly over time, so it can be modeled with low-rank terms, and the moving target satisfies the sparse characteristics in the entire video, so it can be modeled by the coefficient term.
[0004] However, this problem is an NP-hard problem and cannot be solved directly. Based on this, the principal component pursuit model (PCP) uses the nuclear norm and l1 norm to replace the rank function and l0 norm in the corresponding model of formula (1), providing a solid theoretical basis and technical support for the development of the LRSD model. The PCP model has a good video moving target detection effect for simple scenes, but the approximation effect of the rank function and l0 norm used in this method is poor, and it does not consider interference information such as noise, illumination changes and camera jitter, so its video moving target detection effect for complex scenes is poor.
[0005] In recent years, with the rapid development of information technology, people have gradually proposed many improved LRSD models based on the PCP model and have made significant progress. The main idea is to approximate the low-rank terms and sparse terms in equation (1) by introducing appropriate surrogate functions. According to the different surrogate functions, the improved LRSD model can be divided into two categories: convex surrogate function approximation model and non-convex surrogate function approximation model. Summary of the invention
[0006] A video moving object detection method, characterized in that it comprises the following steps:
[0007] Step S1, converting the video to be processed into a two-dimensional matrix with m rows and n columns;
[0008] Step S2, inputting the converted two-dimensional matrix into a pre-built improved robust non-convex norm video moving target detection model;
[0009] Step S3, using an alternating direction multiplier method to solve the improved robust non-convex norm video moving object detection model;
[0010] Step S4, using the video moving target detection model of step S3, solving, extracting and outputting the moving target in the video.
[0011] The invention discloses a video moving target detection method, comprising: converting a video to be processed into a two-dimensional matrix of m rows and n columns; inputting the converted two-dimensional matrix into a pre-constructed video moving target detection model of an improved robust non-convex norm, solving, extracting and outputting moving targets in the video, wherein the pre-constructed video moving target detection model of an improved robust non-convex norm first uses a non-convex Laplace exponential norm to replace the low-rank term in a traditional low-rank sparse decomposition method, secondly uses a non-convex Geman fractional norm to replace the sparse term in a traditional LRSD method, and then introduces a noise term into the model to enhance its anti-noise ability, and finally uses an alternating direction multiplier method to solve the model. The high accuracy of video moving target detection is verified from both visual and quantization perspectives. BRIEF DESCRIPTION OF THE DRAWINGS
[0012] Figure 1 It is a structural block diagram of a video moving target detection method;
[0013] Figure 2 Comparison of visual effects of moving targets extracted using different algorithms. DETAILED DESCRIPTION
[0014] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.
[0015] like Figure 1 As shown, a video moving target detection method includes the following steps:
[0016] Step S1, converting the video to be processed into a two-dimensional matrix with m rows and n columns, where m is the number of frames of the video to be processed, and n is the product of the length and width of each frame of the video.
[0017] Step S2, inputting the converted two-dimensional matrix into a pre-built improved robust non-convex norm video moving object detection model.
[0018] The non-convex Laplace exponential norm and non-convex Geman fractional norm are used to replace the sparsity function and rank function in the traditional low-rank sparse decomposition algorithm, respectively, to achieve a better approximation of the sparse part and low-rank part of the traditional low-rank sparse decomposition algorithm, so as to better extract the moving target from the video; its model is as follows:
[0019]
[0020] stD=B+F+N
[0021] Among them, D is the video observation matrix, B is the low-rank video background matrix, F is the sparse video motion target, N is the noise term, is the Laplace exponential norm for Β, σ t (Β) is the tth singular value of the low-rank matrix B, is the Geman fractional norm for F, F ij is the (i, j)th element of the sparse matrix F, α>0, λ>0, ρ>0 and τ>0 are all balancing factors, is the square of the Frobenius norm of the noise term.
[0022] Step S3, using the alternating direction multiplier method to solve the improved robust non-convex norm video moving target detection model. The specific expression is as follows:
[0023] The augmented Lagrangian function of the corresponding model is:
[0024]
[0025] Where Y is the augmented Lagrange multiplier matrix, μ is the penalty coefficient, and <·,·> is the inner product operation.
[0026] The following is a detailed description of the solution process using step k as an example. The specific steps are as follows:
[0027] First, fix the matrices B, Y, N and the penalty coefficient μ to update F k+1 , the specific solution steps are as follows:
[0028]
[0029] Second, fix the matrices F, Y, N and the penalty coefficient μ to update B k+1 , the specific solution steps are as follows:
[0030]
[0031] in, is the proximity operator of the non-convex function g(·).
[0032] Third, fix the matrices B, F, Y and the penalty coefficient μ to update N k+1 , the specific solution steps are as follows:
[0033]
[0034] Then, use the following equation to solve the above formula, as shown below:
[0035]
[0036] Fourth, update the augmented Lagrange multiplier matrix Y and the penalty coefficient μ, as shown in the following formula:
[0037] Y k+1 =Y k -μ k (B k+1 +F k+1 +N k+1 -D)
[0038] μ k+1 =min(ημ k ,μ max )
[0039] Among them, μ max is the preset maximum value.
[0040] Step S4, using the video moving target detection model of step S3, solving, extracting and outputting the moving target in the video.
[0041] Figure 2 Table 1 is a comparison of the visual effects of moving targets extracted by different algorithms, Table 2 is a comparison of the quantized F-measure values of moving targets extracted by different algorithms, and Table 3 is a comparison of the running time of moving targets extracted by different algorithms. It fully shows that the corresponding moving targets can be extracted by inputting the video into the system, and verifies the effectiveness of the moving target detection system based on fractional norm and minimum and maximum concave penalty proposed by the present invention from both visual and quantitative perspectives.
[0042] Specifically, first, in order to visually verify the superiority of the IRNCN model proposed in the present invention, the present invention randomly selects 691 frames of office, 881 frames of pedestrians, 739 frames of highway, 891 frames of people in shade, 1661 frames of backdoor, 301 frames of park, 1021 frames of skating, 2926 frames of airport, 649 frames of shopping mall, 1285 frames of bootstrap, 1605 frames of water surface, 2424 frames of escalator and 22774 frames of curtain as analysis objects for analysis. Figure 2 As shown. Figure 2 It can be seen that the model using non-convex surrogate function approximation has better video motion target detection effect than the model using convex surrogate function approximation. For example, for office videos, the moving target information extracted by the PCP model and the TNN model is seriously missing, and the information corresponding to the moving target extracted by the Godec model is obviously redundant; for pedestrians videos, the character information extracted by the TNN model and the Godec model has the situation that the background is mistakenly taken as the moving target; for people in shade videos, the moving target extracted by the Godec model has the phenomenon of redundant extracted task information; for skating videos, the skier information extracted by the PCP model is missing, and the Godec model is basically difficult to extract the moving skier information. In comparison, most models using non-convex surrogate function approximation can effectively extract moving targets. The reason for the above phenomenon is that the model using non-convex surrogate function approximation has a good approximation degree to the rank function and norm of the traditional LRSD. In addition, in the model using non-convex surrogate function approximation, the IRNCN proposed in the present invention has better performance than other models. For example, for office videos, skating videos, bootstrap videos and water surface videos, the IRNCN model extracts richer moving target information. For people in shade videos, backdoor videos, water surface videos and curtain videos, the IRNCN model mistakenly identifies fewer backgrounds as moving targets than other models.
[0043] In summary, from a visual perspective, the IRNCN model proposed in this invention has better video moving object detection performance.
[0044] As shown in Table 1, in order to verify the superiority of the IRNCN model from a quantitative perspective, the present invention also gives the F-measure values of video motion target detection of each model. As can be seen from Table 1, whether from the F-measure of a single video or from the average F-measure value, the IRNCN model proposed in the present invention has better performance. For example, the average F-measure value of IRNCN is 0.6500, which is 0.1246, 0.2013, 0.1187, 0.1001, 0.0729, 0.0367 and 0.0329 larger than the PCP model, Godec model, TNN model, GNCR model, NonLRSD model, NLTFN model and NNLSM model, respectively.
[0045] F-measure IRNCN PCP Godec TNN GNCR NonLRSD NLTFN NNLSM airport 0.6092 0.5197 0.5431 0.5656 0.5942 0.5543 0.5991 0.5674 shoppingmall 0.6786 0.6489 0.6470 0.6494 0.6765 0.6529 0.6627 0.6588 Bootstrap 0.5807 0.5601 0.5507 0.5653 0.5664 0.5636 0.5672 0.5726 watersurface 0.7702 0.5440 0.3296 0.4962 0.4862 0.6547 0.7448 0.7569 escalator 0.5704 0.5269 0.5307 0.5416 0.5236 0.5479 0.5562 0.5629 curtain 0.6407 0.4609 0.3861 0.4831 0.5383 0.5026 0.6171 0.6146 backdoor 0.6051 0.5459 0.4308 0.5519 0.5832 0.5784 0.5967 0.5544 office 0.6550 0.4439 0.3616 0.3983 0.2111 0.6329 0.6423 0.6393 car 0.7338 0.6798 0.5138 0.6802 0.6963 0.7069 0.7120 0.7322 park 0.6028 0.5241 0.4291 0.5299 0.5333 0.5398 0.5476 0.5725 pedestrians 0.7475 0.5496 0.3716 0.5495 0.5266 0.5709 0.5714 0.6763 peopleinshade 0.6340 0.4389 0.3791 0.4717 0.6204 0.5644 0.6317 0.5781 Skating 0.6216 0.3868 0.3595 0.4243 0.5925 0.4321 0.5239 0.5358 average value 0.6500 0.5253 0.4487 0.5313 0.5499 0.5770 0.6133 0.6171
[0046] Table 1
[0047] The quantitative data in Table 1 show that the IRNCN model proposed in the present invention has good video moving target detection performance.
[0048] As shown in Table 2, it is a comparison chart of the quantized running time of the moving objects extracted by different algorithms. In order to verify the effectiveness of the IRNCN model in terms of time complexity, Table 2 gives the processing time of each algorithm corresponding to each frame.
[0049] time IRNCN PCP Godec TNN GNCR NonLRSD NLTFN NNLSM airport 0.0348 0.1254 0.0044 0.0690 0.0414 0.0382 0.0360 0.0445 shoppingmall 0.0215 0.0351 0.0049 0.0464 0.0389 0.0374 0.0216 0.0269 Bootstrap 0.0205 0.0380 0.0040 0.0496 0.0332 0.0282 0.0210 0.0264 watersurface 0.0208 0.0418 0.0030 0.0432 0.0312 0.0331 0.0274 0.0315 escalator 0.0268 0.0753 0.0034 0.0524 0.0524 0.0449 0.0453 0.0606 curtain 0.0207 0.0355 0.0049 0.0415 0.0415 0.0312 0.0224 0.0267 backdoor 0.1093 0.2684 0.0151 0.3048 0.1467 0.1985 0.0890 0.1191 office 0.1206 0.2768 0.0170 0.3316 0.1381 0.2512 0.1449 0.1418 car 0.1132 0.2270 0.0148 0.2612 0.1429 0.2390 0.1194 0.1571 park 0.1644 0.5253 0.0421 0.6361 0.3508 0.5109 0.2125 0.2476 pedestrians 0.1224 0.3632 0.0202 0.3525 0.2670 0.2564 0.1757 0.1172 peopleinshade 0.1113 0.2449 0.0181 0.2855 0.1382 0.1823 0.1148 0.1252 Skating 0.1585 0.2227 0.0166 0.2576 0.3245 0.1915 0.1299 0.2962 average value 0.0804 0.1907 0.0130 0.2101 0.1344 0.1571 0.0892 0.1093
[0050] Table 2
[0051] As can be seen from Table 2, the Godec model has the lowest time complexity in video moving target detection because it uses an accelerated bilateral random prediction algorithm. Except for the Godec model, the time complexity of the IRNCN model proposed in the present invention is better than that of other comparison algorithms. For example, the average processing time per frame of the IRNCN model is 0.0804, which is 0.1104s / frame, 0.1297s / frame, 0.0540s / frame, 0.0768s / frame, 0.0089s / frame and 0.0289s / frame less than the average processing time per frame of the PCP model, TNN model, GNCR model, NonLRSD model, NLTFN model and NNLSM model. This verifies that the IRNCN model has good time complexity.
[0052] In summary, the above-mentioned large number of simulation experiments show that the IRNCN model proposed in the present invention has good effectiveness and superiority.
Claims
1. A video moving target detection method, characterized in that: The steps include: Step S1, converting the video to be processed into a two-dimensional matrix with m rows and n columns; Step S2, inputting the converted two-dimensional matrix into a pre-built improved robust non-convex norm video moving target detection model; Step S3, using an alternating direction multiplier method to solve the improved robust non-convex norm video moving object detection model; Step S4, using the video moving target detection model of step S3, solving, extracting and outputting the moving target in the video.
2. A video moving target detection method according to claim 1, characterized in that: In step S1, the video to be processed is converted into a two-dimensional matrix with m rows and n columns, where m is the number of frames of the video to be processed, and n is the product of the length and width of each frame of the video.
3. A video moving target detection method according to claim 2, characterized in that: The improved robust non-convex norm video motion target detection model in step S2 uses the non-convex Laplace exponential norm and the non-convex Geman fractional norm to replace the sparsity function and the rank function in the traditional low-rank sparse decomposition algorithm, respectively, to achieve a better approximation of the sparse part and the low-rank part of the traditional low-rank sparse decomposition algorithm, so as to better extract the motion target from the video; its model is as follows: stD=B+F+N Among them, D is the video observation matrix, B is the low-rank video background matrix, F is the sparse video motion target, N is the noise term, is the Laplace exponential norm for Β, σ t (Β) is the tth singular value of the low-rank matrix B, is the Geman fractional norm for F, F ij is the (i, j)th element of the sparse matrix F, α>0, λ>0, ρ>0 and τ>0 are all balancing factors, is the square of the Frobenius norm of the noise term.
4. A video moving target detection method according to claim 1, characterized in that: The improved robust non-convex norm video moving target detection model in step S3 is solved by using the alternating direction multiplier method, and the specific expression is as follows: The augmented Lagrangian function of the corresponding model is: Where Y is the augmented Lagrange multiplier matrix, μ is the penalty coefficient, and <·,·> is the inner product operation.