Unattended station personnel intrusion identification method and device

Abstract

The application provides a kind of unattended station personnel invasion identification method and device, the method comprises: the picture to be detected is input to personnel invasion identification model, obtains the invasion identification result of personnel invasion identification model output;Wherein, personnel invasion identification model includes image denoising network, feature extraction network, occlusion perception attention network and invasion prediction network, image denoising network is used to denoising sample picture, based on occlusion perception attention network is used to apply sample picture feature, determine the correlation between the occluded human body area and the unoccluded human body area in sample picture.The image denoising network is used to denoise the picture, so as to avoid the noise influence brought by natural environment such as wind, frost, rain, snow and fog to data source.In addition, through the occlusion perception attention network, the problem that human body occlusion caused by the invasion personnel climbing wall affects the invasion personnel identification can be avoided.

Description

Technical Field

[0001] This invention relates to the field of natural gas station technology, and in particular to a method and device for identifying unmanned station personnel intrusion. Background Technology

[0002] With the development of digitalization and intelligentization, unmanned operation and integrated management are inevitable trends for natural gas stations, thanks to relevant remote management technologies. A key characteristic of unmanned stations is the absence of staff on duty. Therefore, ensuring access security and preventing unauthorized entry are the primary issues to be addressed to guarantee the normal operation of unmanned stations.

[0003] Currently, most unmanned stations have perimeter security systems based on vibrating optical fibers. When an intruder causes the optical fiber to vibrate, it alters the internal laser interference phenomenon, which, combined with relevant data processing techniques, is used to determine if the perimeter has been breached. However, surveys of some unmanned stations have revealed that, in actual use, fiber-optic perimeter security systems suffer from high false alarm rates, insufficient stability, and susceptibility to damage. The reliability of fiber-optic perimeter security systems needs further improvement to meet the specific perimeter security requirements of stations. Summary of the Invention

[0004] This invention provides a method and device for identifying unmanned station personnel intrusion, in order to overcome the deficiencies in the prior art.

[0005] This invention provides a method for identifying personnel intrusion at unattended stations, comprising the following steps: Obtain the image to be inspected from the video surveillance system; The image to be detected is input into the intrusion detection model to obtain the intrusion detection result output by the intrusion detection model; The intrusion detection model includes an image denoising network, a feature extraction network, an occlusion awareness attention network, and an intrusion prediction network. The intrusion detection model is trained based on the following steps: The video surveillance system acquires video footage of personnel intrusion at the station perimeter, and the intrusion video is decomposed into sample images frame by frame. The sample images are denoised based on the image denoising network described above. Based on the feature extraction network, sample image features are extracted from the denoised sample images; Based on the occlusion-aware attention network, the correlation between occluded and unoccluded human body regions in the sample image is determined by applying the features of the sample image. Based on the intrusion prediction network, the correlation and sample image features are applied to determine the current human body prediction box, the human body ground box and other human body prediction boxes. The human body ground box refers to the human detection label that has the greatest overlap with the current human body prediction box, excluding the current human body prediction box. Based on the current human body prediction bounding box and the human body real bounding box, the first occlusion perception rejection loss is determined; Based on the current human body prediction box and the other human body prediction boxes, determine the second occlusion perception rejection loss; Based on the first occlusion perception rejection loss and the second occlusion perception rejection loss, the parameters of the personnel intrusion recognition model are updated to obtain the trained personnel intrusion recognition model.

[0006] According to the present invention, an unmanned station personnel intrusion identification method is provided, wherein the image denoising network includes an estimation module and a denoised image generation module; The denoising of the sample image based on the image denoising network includes: Based on the estimation module, the sample image is used to determine the estimated depth and relative noise level; Based on the denoised image generation module, the estimated depth and the relative noise level are applied to generate denoised sample images.

[0007] According to the present invention, an unattended station personnel intrusion identification method is provided, wherein determining a first occlusion perception rejection loss based on the current human body prediction bounding box and the human body real bounding box includes: Based on the overlap between the current human body prediction bounding box and the human body real bounding box, the true intersection with the human body real annotation box is determined; Based on the actual intersection with the actual human body bounding box, the first occlusion perception rejection loss is determined.

[0008] According to the present invention, in an unattended station personnel intrusion identification method, the second occlusion perception rejection loss is determined based on the overlap between the current human prediction box and the other human prediction boxes.

[0009] According to the unmanned station intrusion identification method provided by the present invention, after obtaining the trained intrusion identification model, it further includes: Obtain the test image and the corresponding test label for the test image; The test image is input into the intrusion detection model to obtain the test prediction result output by the intrusion detection model. Based on the test prediction results and the test labels, an confusion matrix is ​​established; Based on the confusion matrix, the evaluation index of the personnel intrusion identification model is determined; Based on the evaluation metrics, the recognition performance of the personnel intrusion recognition model is determined.

[0010] According to the present invention, an unattended station personnel intrusion identification method is provided, wherein establishing a confusion matrix based on the test prediction results and the test labels includes: Based on the test prediction results and the test labels, true positives, true negatives, false positives, and false negatives are determined. The confusion matrix is ​​established based on the true positive examples, the true negative examples, the false positive examples, the false negative examples, the positive examples in the test prediction results, and the negative examples in the test prediction results.

[0011] According to the present invention, an unmanned station personnel intrusion identification method is provided, wherein the evaluation indicators include precision, recall, and mean precision.

[0012] The present invention also provides an unmanned station personnel intrusion identification device, comprising the following modules: The acquisition unit is used to acquire images to be detected from the video surveillance system. The recognition unit is used to input the image to be detected into the personnel intrusion recognition model and obtain the intrusion recognition result output by the personnel intrusion recognition model; The intrusion detection model includes an image denoising network, a feature extraction network, an occlusion awareness attention network, and an intrusion prediction network. The intrusion detection model is trained based on the following steps: The video surveillance system acquires video footage of personnel intrusion at the station perimeter, and the intrusion video is decomposed into sample images frame by frame. The sample images are denoised based on the image denoising network described above. Based on the feature extraction network, sample image features are extracted from the denoised sample images; Based on the occlusion-aware attention network, the correlation between occluded and unoccluded human body regions in the sample image is determined by applying the features of the sample image. Based on the intrusion prediction network, the correlation and sample image features are applied to determine the current human body prediction box, the human body ground box and other human body prediction boxes. The human body ground box refers to the human detection label that has the greatest overlap with the current human body prediction box, excluding the current human body prediction box. Based on the current human body prediction bounding box and the human body real bounding box, the first occlusion perception rejection loss is determined; Based on the current human body prediction box and the other human body prediction boxes, determine the second occlusion perception rejection loss; Based on the first occlusion perception rejection loss and the second occlusion perception rejection loss, the parameters of the personnel intrusion recognition model are updated to obtain the trained personnel intrusion recognition model.

[0013] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement any of the above-described unattended station personnel intrusion identification methods.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the unattended station personnel intrusion identification method as described above.

[0015] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the unmanned station personnel intrusion identification method described above.

[0016] The unmanned station intrusion identification method and device provided by this invention uses an image denoising network to denoise images, thereby avoiding the noise impact of natural environments such as wind, frost, rain, snow, and fog on the data source. Furthermore, embodiments of this invention utilize an occlusion-sensing attention network to avoid the problem of intruders being obstructed by their bodies when climbing over walls, thus affecting the identification of intruders. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the unmanned station personnel intrusion identification method provided by the present invention.

[0019] Figure 2 This is a schematic diagram of the structure of the personnel intrusion identification model provided by the present invention.

[0020] Figure 3 This is a schematic diagram of the unmanned station personnel intrusion identification device provided by the present invention.

[0021] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0023] Figure 1 This is a flowchart illustrating the unmanned station personnel intrusion identification method provided by the present invention, as shown below. Figure 1 As shown, the method includes steps 110 and 120.

[0024] Step 110: Obtain the image to be detected from the video surveillance system; Step 120: Input the image to be detected into the personnel intrusion detection model to obtain the intrusion detection result output by the personnel intrusion detection model; The intrusion detection model includes an image denoising network, a feature extraction network, an occlusion awareness attention network, and an intrusion prediction network. The intrusion detection model is trained based on the following steps: The video surveillance system was used to obtain video footage of personnel intrusion at the perimeter of the station, and the intrusion video was decomposed into sample images frame by frame. Based on an image denoising network, denoise the sample images; Based on the feature extraction network, the features of the sample image are extracted from the denoised sample image. Based on the occlusion-aware attention network, the correlation between occluded and unoccluded human body regions in the sample image is determined by applying sample image features. Based on the intrusion prediction network, the correlation and sample image features are applied to determine the current human body prediction box, the human body ground box and other human body prediction boxes. The human body ground box refers to the human detection label that has the greatest overlap with the current human body prediction box, excluding the current human body prediction box. Based on the current human body prediction bounding box and the actual human body bounding box, determine the first occlusion perception rejection loss; Based on the current human body prediction box and other human body prediction boxes, determine the second occlusion perception rejection loss; Based on the first occlusion perception rejection loss and the second occlusion perception rejection loss, the parameters of the personnel intrusion recognition model are updated to obtain the trained personnel intrusion recognition model.

[0025] Specifically, in this embodiment of the invention, a person intrusion recognition model is first trained, and then the intrusion recognition result is determined based on the image to be detected using the person intrusion recognition model. The person intrusion recognition model includes an image denoising network, a feature extraction network, an occlusion awareness attention network, and an intrusion prediction network. The image denoising network is used to remove noise from the sample image, and the feature extraction network is used to extract sample image features from the noise-removed sample image. The occlusion awareness attention network is used to apply the sample image features to determine the correlation between occluded and unoccluded human body regions in the sample image. The intrusion prediction network is used to apply the correlation and sample image features to determine the current human body prediction bounding box, the ground truth human body bounding box, and other human body prediction bounding boxes.

[0026] Furthermore, the intrusion detection model can be obtained by improving the YOLOv8 network. Specifically, the AODNet network is inserted as the image denoising network into the head of the YOLOv8 network, and the SEAM module is inserted as the occlusion perception attention network into the neck of the YOLOv8 network. This can solve the problem of human occlusion and noise affecting the identification of intruders.

[0027] Specifically, the SEAM attention module combines depthwise separable convolutions with residual connections to compensate for the response loss of the occluded part of the detected object and enhance the response of the unoccluded part. The SEAM architecture mainly consists of two parts: an occlusion-aware attention network and an occlusion-aware rejection loss (i.e., the first occlusion-aware rejection loss and the second occlusion-aware rejection loss).

[0028] The SEAM module has three main objectives: to facilitate multi-scale human detection, to highlight human regions in images, and conversely, to weaken background regions. The first part of SEAM is a depthwise separable convolution with residual connections. Depthwise separable convolution operates on a depthwise basis, meaning the convolution separates channels one by one. While depthwise separable convolution can learn the importance of different channels and reduce the number of parameters, it ignores the information relationships between channels. To compensate for this loss, the outputs of convolutions at different depths are subsequently combined through pointwise (1×1) convolutions. Then, two fully connected layers aggregate information from each channel, enhancing the inter-channel connections. This framework aims to compensate for the information loss caused by occlusion by utilizing the learned relationships between occluded and unoccluded human regions. An exponential transformation is performed on the logarithms obtained from the fully connected layers, expanding the value range from [0,1] to [1,e]. This exponential normalization provides a monotonic mapping, making the results more tolerant of positional errors. Finally, the output of the SEAM module is multiplied by the original features, enabling the model to effectively address object occlusion problems.

[0029] The unmanned station intrusion identification method provided in this invention uses an image denoising network to denoise images, thereby avoiding the noise impact of natural environments such as wind, frost, rain, snow, and fog on the data source. Furthermore, this invention utilizes an occlusion-sensing attention network to avoid the problem of intruders being obstructed by their bodies when climbing over walls, thus affecting the identification of intruders.

[0030] Based on the above embodiments, the image denoising network includes an estimation module and a denoised image generation module; Based on an image denoising network, denoising is performed on sample images, including: Based on the estimation module, sample images are used to determine the estimated depth and relative noise level; Based on the denoising image generation module, the depth is estimated and the relative noise level is used to generate denoised sample images.

[0031] As an optional implementation, the image denoising network can be built based on the AODNet network (All-in-One Dehazing Network, a deep learning network for image dehazing). The AODNet network includes a K estimation module and a denoised image generation module.

[0032] The K-estimation module is a key component of the AODNet network, responsible for estimating depth and relative noise levels. It consists of five convolutional layers that fuse filters of different sizes to form multi-scale features. Each convolutional layer uses only three filters, making the AODNet network structure very lightweight. Following the K-estimation module, the denoising image generation module comprises one element-wise multiplication layer and several element-wise addition layers to generate the restored image.

[0033] The denoised image generation module calculates the denoised sample image by taking the result K(x) from the K estimation module. The denoising formula is shown below: J(x)=K(x)I(x)-K(x)+b

[0034] Where I(x) is the observed original sample image, and J(x) is the denoised sample image. A represents global atmospheric light, and t(x) is the transmission matrix, defined as: t(x)=e -βd(x)

[0035] Where β is the atmospheric scattering coefficient, and d(x) is the distance between the object and the camera in the sample image.

[0036] In this way, 1 / t(x) and A are both integrated into the new variable K(x). b is a constant deviation with a default value of 1. Since K(x) depends on I(x), the goal of this embodiment is to establish an input-adaptive depth model whose parameters will vary with the input blurred image, thereby minimizing the reconstruction error between the output J(x) and the clean ground image.

[0037] Based on any of the above embodiments, the first occlusion perception rejection loss is determined based on the current human body prediction bounding box and the human body ground truth bounding box, including: Based on the overlap between the current human body prediction bounding box and the actual human body bounding box, determine the actual intersection with the actual human body annotation box; Based on the actual intersection with the actual human body bounding box, the first occlusion perception rejection loss is determined.

[0038] The second occlusion perception rejection loss is determined based on the overlap between the current human body prediction box and other human body prediction boxes.

[0039] Specifically, when there are multiple intruders in a sample image, intra-class occlusion may cause human body A to contain features of human body B, leading to a high false detection rate. Introducing the first occlusion-aware rejection loss RepGT and the second occlusion-aware rejection loss RepBox can effectively alleviate this problem through rejection. The function of the first occlusion-aware rejection loss RepGT is to keep the current human body prediction box as far away as possible from the surrounding ground truth human body boxes. Here, the ground truth human body box refers to the human detection label with the maximum value other than the object returned by the current human body prediction box itself. The formula for calculating the first occlusion-aware rejection loss RepGT is as follows:

[0040] Wherein, in the formula

[0041] Where P is the current predicted human bounding box, and Re is the ground truth human bounding box with the maximum IoU.

[0042] The overlap between P and Re is defined as the intersection with the actual body bounding box (IoG): IoG(P, G) = area(P∩G) / area(G), IoG(B, G) ∈ [0, 1]. ln It is continuously differentiable in the range (0,1). σ∈[0,1) is a smoothing parameter used to adjust the sensitivity of the repulsive force loss to outliers.

[0043] Furthermore, the purpose of the second occlusion-aware repulsion loss (RepBox) is to keep the current human prediction box as far away from other human prediction boxes as possible, reducing the IoU between them and preventing two different human bodies from being predicted as the same human body. The formula for calculating the second occlusion-aware repulsion loss (RepBox) is as follows:

[0044] Among them, P i P is the current human body prediction bounding box. j For other human prediction boxes, I represents the identity function, and ε represents a small constant when divided by zero.

[0045] As an optional embodiment, the structure of the above-mentioned personnel intrusion identification model can be obtained by improving the original YOLOv8 network using the AODNet network and the SEAM module. Figure 2 This is a schematic diagram of the personnel intrusion identification model provided by the present invention, as shown below. Figure 2 As shown, the AODNet network optimizes the backbone of the original YOLOv8 network. Essentially, the AODNet network acts as an image denoising network, first denoising the sample images to eliminate the impact of noise on recognition, and then extracting features from the denoised sample images. The SEAM module acts as an occlusion-aware attention network, optimizing the neck region of the original YOLOv8 network. By fusing the SEAM module into the neck region, the fully connected networks of the preceding and following layers can aggregate information from each channel, enhancing inter-channel connectivity. Information loss caused by occlusion is compensated by utilizing the learned relationships between occluded and uncrowded human body regions.

[0046] Based on any of the above embodiments, after obtaining the trained personnel intrusion detection model, the method further includes: Obtain the test image and its corresponding test tag; Input the test image into the intrusion detection model to obtain the test prediction results output by the intrusion detection model; Based on the test prediction results and test labels, an confusion matrix is ​​established; Based on the confusion matrix, the evaluation index of the personnel intrusion identification model is determined; The recognition performance of the human intrusion detection model is determined based on evaluation metrics.

[0047] Specifically, test images can be obtained based on the following steps: Images of individuals attempting to bypass internet restrictions and gain access were collected. These images were then labeled using the make-sense software, creating a dataset with a one-to-one correspondence between images and labels. 80% of the images were randomly selected as sample images, and the remaining 20% ​​were used as test images.

[0048] The confusion matrix is ​​constructed based on the test prediction results and test labels, including: Based on the test prediction results and test labels, identify true positives, true negatives, false positives, and false negatives; A confusion matrix is ​​constructed based on true positives, true negatives, false positives, false negatives, positives in the test prediction results, and negatives in the test prediction results.

[0049] To reasonably evaluate the performance of the intrusion detection model, a confusion matrix is ​​first established. This confusion matrix is ​​constructed based on true positives, true negatives, false positives, false negatives, positive examples from the test prediction results, and negative examples from the test prediction results. A true positive (TP) indicates a positive test prediction result that matches the test label. A true negative (TN) indicates a negative test prediction result that matches the test label. A false positive (FP) indicates a positive test prediction result that differs from the test label. A false negative (FN) indicates a negative test prediction result that differs from the test label. Table 1 shows the confusion matrix list provided by this invention. Figure 1 As shown, when both the test prediction result and the test label are positive, it indicates a true example; when the test prediction result is positive and the test label is negative, it indicates a false negative; when both the test result and the test label are negative, it indicates a true negative; and when the test prediction result is positive and the test label is negative, it indicates a false positive.

[0050] Table 1

[0051] Based on any of the above embodiments, the evaluation metrics include precision, recall, and mean precision (mAP). Precision (P), recall (R), and mean precision (mAP) can be determined based on the following formula:

[0052] in,

[0053] The unmanned station personnel intrusion identification device provided by the present invention is described below. The unmanned station personnel intrusion identification device described below can be referred to in correspondence with the unmanned station personnel intrusion identification method described above.

[0054] Based on any of the above embodiments Figure 3 This is a schematic diagram of the unmanned station personnel intrusion identification device provided by the present invention, as shown below. Figure 3 As shown, the device includes: The acquisition unit 310 is used to acquire the image to be detected from the video surveillance system. The recognition unit 320 is used to input the image to be detected into the personnel intrusion recognition model and obtain the intrusion recognition result output by the personnel intrusion recognition model; The intrusion detection model includes an image denoising network, a feature extraction network, an occlusion awareness attention network, and an intrusion prediction network. The intrusion detection model is trained based on the following steps: The video surveillance system was used to obtain video footage of personnel intrusion at the perimeter of the station, and the intrusion video was decomposed into sample images frame by frame. Based on an image denoising network, denoise the sample images; Based on the feature extraction network, the features of the sample image are extracted from the denoised sample image. Based on the occlusion-aware attention network, the correlation between occluded and unoccluded human body regions in the sample image is determined by applying sample image features. Based on the intrusion prediction network, the correlation and sample image features are applied to determine the current human body prediction box, the human body ground box and other human body prediction boxes. The human body ground box refers to the human detection label that has the greatest overlap with the current human body prediction box, excluding the current human body prediction box. Based on the current human body prediction bounding box and the actual human body bounding box, determine the first occlusion perception rejection loss; Based on the current human body prediction box and other human body prediction boxes, determine the second occlusion perception rejection loss; Based on the first occlusion perception rejection loss and the second occlusion perception rejection loss, the parameters of the personnel intrusion recognition model are updated to obtain the trained personnel intrusion recognition model.

[0055] Image denoising networks include an estimation module and a denoised image generation module; Based on an image denoising network, denoising is performed on sample images, including: Based on the estimation module, sample images are used to determine the estimated depth and relative noise level; Based on the denoising image generation module, the depth is estimated and the relative noise level is used to generate denoised sample images.

[0056] Based on any of the above embodiments, the first occlusion perception rejection loss is determined based on the current human body prediction bounding box and the human body ground truth bounding box, including: Based on the overlap between the current human body prediction bounding box and the actual human body bounding box, determine the actual intersection with the actual human body annotation box; Based on the actual intersection with the actual human body bounding box, the first occlusion perception rejection loss is determined.

[0057] Based on any of the above embodiments, the second occlusion perception rejection loss is determined based on the overlap between the current human prediction box and other human prediction boxes.

[0058] Based on any of the above embodiments, after obtaining the trained personnel intrusion detection model, the method further includes: Obtain the test image and its corresponding test tag; Input the test image into the intrusion detection model to obtain the test prediction results output by the intrusion detection model; Based on the test prediction results and test labels, an confusion matrix is ​​established; Based on the confusion matrix, the evaluation index of the personnel intrusion identification model is determined; The recognition performance of the human intrusion detection model is determined based on evaluation metrics.

[0059] Based on any of the above embodiments, based on the test prediction results and test labels, a confusion matrix is ​​established, including: based on the test prediction results and test labels, determining true positives, true negatives, false positives and false negatives; A confusion matrix is ​​constructed based on true positives, true negatives, false positives, false negatives, positives in the test prediction results, and negatives in the test prediction results.

[0060] Based on any of the above embodiments, the evaluation metrics include precision, recall, and mean precision.

[0061] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 4 As shown, the electronic device may include a processor 410, a communications interface 420, a memory 430, and a communication bus 440. The processor 410, communications interface 420, and memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions from the memory 430 to execute an unattended site personnel intrusion detection method.

[0062] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0063] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the unattended station personnel intrusion identification method provided by the above methods.

[0064] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the unattended station personnel intrusion identification method provided by the above methods.

[0065] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0066] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying personnel intrusion at an unattended station, characterized in that, include: Obtain the image to be inspected from the video surveillance system; The image to be detected is input into the intrusion detection model to obtain the intrusion detection result output by the intrusion detection model; The intrusion detection model includes an image denoising network, a feature extraction network, an occlusion awareness attention network, and an intrusion prediction network. The intrusion detection model is trained based on the following steps: The video surveillance system acquires video footage of personnel intrusion at the station perimeter, and the intrusion video is decomposed into sample images frame by frame. The sample images are denoised based on the image denoising network described above. Based on the feature extraction network, sample image features are extracted from the denoised sample images; Based on the occlusion-aware attention network, the correlation between occluded and unoccluded human body regions in the sample image is determined by applying the features of the sample image. Based on the intrusion prediction network, the correlation and sample image features are applied to determine the current human body prediction box, the human body ground box and other human body prediction boxes. The human body ground box refers to the human detection label that has the greatest overlap with the current human body prediction box, excluding the current human body prediction box. Based on the current human body prediction bounding box and the human body real bounding box, the first occlusion perception rejection loss is determined; Based on the current human body prediction box and the other human body prediction boxes, determine the second occlusion perception rejection loss; Based on the first occlusion perception rejection loss and the second occlusion perception rejection loss, the parameters of the personnel intrusion recognition model are updated to obtain the trained personnel intrusion recognition model.

2. The unmanned station personnel intrusion identification method according to claim 1, characterized in that, The image denoising network includes an estimation module and a denoised image generation module; The denoising of the sample image based on the image denoising network includes: Based on the estimation module, the sample image is used to determine the estimated depth and relative noise level; Based on the denoised image generation module, the estimated depth and the relative noise level are applied to generate denoised sample images.

3. The unmanned station personnel intrusion identification method according to claim 1, characterized in that, The step of determining the first occlusion perception rejection loss based on the current human body prediction bounding box and the human body ground truth bounding box includes: Based on the overlap between the current human body prediction bounding box and the human body real bounding box, the true intersection with the human body real annotation box is determined; Based on the actual intersection with the actual human body bounding box, the first occlusion perception rejection loss is determined.

4. The unmanned station personnel intrusion identification method according to claim 1, characterized in that, The second occlusion perception rejection loss is determined based on the overlap between the current human prediction box and the other human prediction boxes.

5. The unmanned station personnel intrusion identification method according to claim 1, characterized in that, After obtaining the trained intrusion detection model, the following is also included: Obtain the test image and the corresponding test label for the test image; The test image is input into the intrusion detection model to obtain the test prediction result output by the intrusion detection model. Based on the test prediction results and the test labels, an confusion matrix is ​​established; Based on the confusion matrix, the evaluation index of the personnel intrusion identification model is determined; Based on the evaluation metrics, the recognition performance of the personnel intrusion recognition model is determined.

6. The unmanned station personnel intrusion identification method according to claim 5, characterized in that, The step of establishing a confusion matrix based on the test prediction results and the test labels includes: Based on the test prediction results and the test labels, true positives, true negatives, false positives, and false negatives are determined. The confusion matrix is ​​established based on the true positive examples, the true negative examples, the false positive examples, the false negative examples, the positive examples in the test prediction results, and the negative examples in the test prediction results.

7. The unmanned station personnel intrusion identification method according to claim 5, characterized in that, The evaluation metrics include precision, recall, and mean precision.

8. An unmanned station personnel intrusion identification device, characterized in that, include: The acquisition unit is used to acquire images to be detected from the video surveillance system. The recognition unit is used to input the image to be detected into the personnel intrusion recognition model and obtain the intrusion recognition result output by the personnel intrusion recognition model; The intrusion detection model includes an image denoising network, a feature extraction network, an occlusion awareness attention network, and an intrusion prediction network. The intrusion detection model is trained based on the following steps: The video surveillance system acquires video footage of personnel intrusion at the station perimeter, and the intrusion video is decomposed into sample images frame by frame. The sample images are denoised based on the image denoising network described above. Based on the feature extraction network, sample image features are extracted from the denoised sample images; Based on the occlusion-aware attention network, the correlation between occluded and unoccluded human body regions in the sample image is determined by applying the features of the sample image. Based on the intrusion prediction network, the correlation and sample image features are applied to determine the current human body prediction box, the human body ground box and other human body prediction boxes. The human body ground box refers to the human detection label that has the greatest overlap with the current human body prediction box, excluding the current human body prediction box. Based on the current human body prediction bounding box and the human body real bounding box, the first occlusion perception rejection loss is determined; Based on the current human body prediction box and the other human body prediction boxes, determine the second occlusion perception rejection loss; Based on the first occlusion perception rejection loss and the second occlusion perception rejection loss, the parameters of the personnel intrusion recognition model are updated to obtain the trained personnel intrusion recognition model.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the unmanned station personnel intrusion identification method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the unmanned station personnel intrusion identification method as described in any one of claims 1 to 7.

Images