Intelligent recognition method and equipment for important parts of hidden target

Abstract

The invention provides an intelligent recognition method and device for important parts of a hidden target, and relates to the field of hidden target recognition, and the method comprises the steps: constructing a real-time shielding target recognition model, training the real-time shielding target recognition model through a training image, and obtaining a trained real-time shielding target recognition model; obtaining an image of a to-be-identified area and performing preprocessing to generate a preprocessed image; inputting the preprocessed image into a trained real-time shielding target recognition model to obtain a comprehensive final whole body anchor point and a comprehensive whole body classification result; and analyzing the comprehensive final whole body anchor point and the comprehensive whole body classification result to obtain important parts of the target. The visible target positioning network and the invisible part positioning network can respectively obtain visible part anchor points and whole body anchor points in the image. And the important part positioning association network makes up the gap between the visible part anchor points and the whole body anchor points by using block statistics, so that the recognition precision of the important part of the target is improved.

Description

Technical Field

[0001] The present invention relates to the field of concealed target recognition, and in particular to a method and device for intelligently recognizing key parts of a concealed target. Background Art

[0002] Target detection is an important task in the field of computer vision. Its main purpose is to detect targets (objects) in images or videos and determine the category and location of the targets. It is an important part of scene understanding. Military target detection plays a very important role in modern warfare. Guided weapons with target detection capabilities can ensure accurate and effective strikes and destruction of targets, and can also detect enemy situations in a timely manner to reduce casualties. This requires weapon systems to have fast and efficient detection and recognition capabilities. With the rapid development of deep learning technology and the improvement of computing hardware performance, target detection algorithms have made great progress and have strong accuracy and generalization capabilities.

[0003] The key parts refer to the positions or parts that are critical and important in a system or structure. These parts usually play a vital role in the overall function, structural strength or safety of the system. Once these parts are damaged or fail, the function of the entire system may be seriously affected, and even cause the system to collapse or fail. The key parts of the human body mainly include the head, chest, abdomen, spine and limbs. The head is the control center of the human body, the chest contains important organs such as the heart and lungs, the abdomen protects the liver, stomach, kidneys and other organs necessary for life, and the spine transmits nerve signals between the brain and the whole body. In addition, joints and limbs are also very important for motor function. The key parts of a vehicle include the engine, brake system, tires, steering system, fuel system, suspension system and body structure. The engine provides the power source, the brake system and tires are related to safety, the steering system controls the direction of the vehicle, the fuel system ensures normal fuel supply, and the suspension system and body structure ensure the stability and safety of driving. Whether it is the human body or the vehicle, damage or failure of key parts may lead to serious consequences, and special attention should be paid to these parts during detection and monitoring. Therefore, it is of great significance to study the intelligent recognition method of key parts of concealed targets and effectively identify and detect key targets and strike weak targets in military environments.

[0004] However, when ordinary target detection algorithms are directly applied to military battlefield environments, their accuracy and generalization ability will be greatly reduced, not only because of the harsh extreme weather in battlefield environments and the diversity of target detection application scenarios that are more diverse than in daily life, but also because interference is more frequent and prominent in military scenarios. Different from general scenes, there will be occlusion interference such as smoke bombs and decoy bombs. Occlusion conditions refer to the situation where the target is blocked by other objects or targets of the same type, which is particularly common in military environments. Occlusion may cause part or all of the target to be hidden, making it difficult for the detection system to determine the complete shape and position of the target. In addition, occlusion will blur the boundary between the target and the background, increasing the possibility of false detection and false alarm. In military strikes, occlusion situations may be more complex and common. For example, in urban battlefields, various obstacles such as buildings, vehicles and crowds will interfere with and make it difficult to identify targets.

[0005] Therefore, there is an urgent need for an intelligent recognition system that can ensure high recognition accuracy and good real-time performance, so as to realize real-time monitoring and early warning of occluded targets. Summary of the invention

[0006] In view of this, the purpose of the present invention is to provide a method and device for intelligently identifying the key parts of a hidden target, so as to solve the problem that the recognition accuracy of the key parts of the hidden target is not high.

[0007] The present invention provides a method for intelligently identifying key parts of a concealed target, comprising the following steps:

[0008] S1: construct a real-time occluded target recognition model, train the real-time occluded target recognition model through training images, and obtain a trained real-time occluded target recognition model;

[0009] S2: Acquire the image of the area to be identified and perform preprocessing to generate a preprocessed image;

[0010] S3: Input the preprocessed image into the trained real-time occluded object recognition model to obtain the comprehensive final full-body anchor point and comprehensive full-body classification results;

[0011] S4: Analyze the comprehensive final whole-body anchor points and the comprehensive whole-body classification results to obtain the key parts of the target.

[0012] Preferred:

[0013] The real-time occluded target recognition model includes: a single-stage recognizer, a first recognition channel, a second recognition channel, a third recognition channel, a fourth recognition channel, a first receptive field module, a second receptive field module and a third receptive field module;

[0014] The single-stage identifier is connected to the first identification channel, the second identification channel, the third identification channel and the fourth identification channel;

[0015] The output end of the fourth identification channel, the third receptive field module, the output end of the third identification channel, the second receptive field module, the output end of the second identification channel, the first receptive field module and the output end of the first identification channel are connected in sequence.

[0016] Preferred:

[0017] The first identification channel, the second identification channel, the third identification channel and the fourth identification channel have the same structure;

[0018] The first recognition channel includes a visible target positioning network, a key part positioning association network and an invisible part positioning network connected in sequence;

[0019] The visible target positioning network includes: a first backbone network and a first recognition head;

[0020] The key part positioning association network includes: a second backbone network and a key part identification network;

[0021] The invisible part localization network includes: a third backbone network and a second recognition head;

[0022] The structures of the first backbone network, the second backbone network and the third backbone network are the same;

[0023] The first backbone network includes an input layer, a first bottleneck layer, a second bottleneck layer, a third bottleneck layer, a fourth bottleneck layer and an output layer which are connected in sequence.

[0024] Preferably, step S1 specifically comprises:

[0025] S11: input the training image into the single-stage recognizer to obtain a first size feature map, a second size feature map, a third size feature map, and a fourth size feature map;

[0026] S12: The first recognition channel, the second recognition channel, the third recognition channel and the fourth recognition channel are trained respectively through the first size feature map, the second size feature map, the third size feature map and the fourth size feature map to obtain a trained real-time occluded target recognition model.

[0027] Preferably, the training processes of the first recognition channel, the second recognition channel, the third recognition channel and the fourth recognition channel are the same, and the training process of the first recognition channel is specifically as follows:

[0028] S121: Set a pre-anchor point A0, input the first size feature map into the first backbone network to obtain a visible part image, input the visible part image and the pre-anchor point A0 into the first recognition head, obtain a visible part anchor point A1 and a visible classification result, and calculate the loss L of the visible classification result VE ;

[0029] S122: Repeat step S121 until the loss is L VE Less than the preset value, a trained visible target positioning network is obtained;

[0030] S123: Divide the human body into three scales, including: key point scale s1, low-level local scale s2 and high-level local scale s3, and calculate the semantic local feature V of the key point scale s1 s1 and the global feature V g ;

[0031] S124: By using the semantic local feature V s1 Perform an average pooling operation to calculate the semantic local features V of the low-level local scale s2 s2 and the semantic local features V of the high-level local scale s3 s3 ;

[0032] S125: Through semantic local features V s1 、V s2 、V s3 and the global feature V g The visible part of the image is segmented into human bodies and input into the key part positioning association network to obtain the initial whole body anchor point A2;

[0033] S126: Input the first size feature map into the third backbone network to obtain a full-body image, input the full-body image and the initial full-body anchor point A2 into the second recognition head, obtain the final full-body anchor point A3 and the full-body classification result, and calculate the loss L of the full-body classification result FR ;

[0034] S127: Repeat step S126 until the loss is L FR is less than a preset value, a trained invisible part positioning network is obtained, and a trained first recognition channel is output.

[0035] Preferred:

[0036] Loss L VE The calculation formula is:

[0037] L VE =L local +λ v [y=1]L smoothL1

[0038] Among them, L local is the focal loss of the visible classification result, L smoothL1 is the smooth-L1 loss of A1 regression, [y=1] is the indicator of positive samples, λ v is the first tuning parameter.

[0039] Preferred:

[0040] Semantic local feature V s1 and the global feature V g The calculation formula is:

[0041]

[0042] V g =g GAP (F)

[0043] in, and g GAP They refer to the outer product and global average pooling operations respectively; F is the feature map obtained after the visible part image is input into the second backbone network, It is a heat map of key parts of the human body obtained by the key part recognition network.

[0044] Preferred:

[0045] Loss L FR The calculation formula is:

[0046] L FR =L local +λ f [y=1]L occ

[0047]

[0048] Among them, L local is the focal loss of the visible classification result, [y=1] is the indicator of positive samples, λ f is the second tuning parameter; L occ is the weighted loss, a is the anchor box of the initial full-body anchor A2, b∈B gt , B gt is the set of true bounding boxes, s is the difference between the predicted offset and the true ground-truth offset, and IoU(a,b) is the intersection over union (IoU) between the anchor box of the initial full-body anchor and the true bounding box.

[0049] Preferably, step S3 specifically comprises:

[0050] S31: Input the preprocessed image into the trained real-time occluded target recognition model, and obtain the first final whole-body anchor point, the first whole-body classification result, the second final whole-body anchor point, the second whole-body classification result, the third final whole-body anchor point, the third whole-body classification result, the fourth final whole-body anchor point, and the fourth whole-body classification result through the first recognition channel, the second recognition channel, the third recognition channel, and the fourth recognition channel respectively;

[0051] S32: adjusting the first final whole-body anchor point and the first whole-body classification result through the first receptive field module, adjusting the second final whole-body anchor point and the second whole-body classification result through the second receptive field module, and adjusting the third final whole-body anchor point and the third whole-body classification result through the third receptive field module;

[0052] S33: Fuse the first final whole-body anchor point, the second final whole-body anchor point, the third final whole-body anchor point and the fourth final whole-body anchor point to obtain a comprehensive final whole-body anchor point; fuse the first whole-body classification result, the second whole-body classification result, the third whole-body classification result and the fourth whole-body classification result to obtain a comprehensive whole-body classification result.

[0053] A device for intelligently identifying key parts of a concealed target comprises: a processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement the method for intelligently identifying key parts of a concealed target.

[0054] The present invention has the following beneficial effects:

[0055] The recognition channel in the real-time occluded target recognition model of the present invention is composed of a visible target positioning network, a key part positioning association network and an invisible part positioning network. The visible target positioning network and the invisible part positioning network can respectively obtain the visible part anchor point and the whole body anchor point in the image.

[0056] The visible object localization network and the invisible part localization network use two independent recognition heads to complete full-body pedestrian recognition. This architecture achieves progressive design and improves recognition effect without significantly increasing complexity through two independent recognition heads.

[0057] The key parts localization association network uses occlusion statistics to make up for the gap between visible part anchors and full body anchors, and introduces a receptive field module to provide a diverse receptive field, optimize the recognition of pedestrians of various scales, and improve the recognition accuracy of key parts of the target. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a flow chart of a method according to an embodiment of the present invention;

[0059] Figure 2 This is a structural diagram of the real-time occluded target recognition model;

[0060] Figure 3 This is the structural diagram of the recognition head and receptive field module;

[0061] Figure 4 This is the structural diagram of the backbone network;

[0062] Figure 5 This is a structural diagram of the device according to an embodiment of the present invention;

[0063] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0064] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0065] Reference Figure 1 The present invention provides a method for intelligently identifying key parts of a concealed target, which is characterized by comprising the following steps:

[0066] S1: construct a real-time occluded target recognition model, train the real-time occluded target recognition model through training images, and obtain a trained real-time occluded target recognition model;

[0067] Furthermore, the real-time occluded target recognition model is built based on the Progressive Refinement Network (PRNet) framework, such as Figure 2 shown.

[0068] The real-time occluded target recognition model includes: a single-stage recognizer, a first recognition channel, a second recognition channel, a third recognition channel, a fourth recognition channel, a first receptive field module, a second receptive field module and a third receptive field module;

[0069] The single-stage identifier is connected to the first identification channel, the second identification channel, the third identification channel and the fourth identification channel;

[0070] The output end of the fourth identification channel, the third receptive field module, the output end of the third identification channel, the second receptive field module, the output end of the second identification channel, the first receptive field module and the output end of the first identification channel are connected in sequence.

[0071] Further, the first identification channel, the second identification channel, the third identification channel and the fourth identification channel have the same structure;

[0072] The first recognition channel includes a visible target positioning network, a key part positioning association network and an invisible part positioning network connected in sequence;

[0073] The visible target positioning network includes: a first backbone network and a first recognition head;

[0074] The key part positioning association network includes: a second backbone network and a key part identification network;

[0075] The invisible part localization network includes: a third backbone network and a second recognition head;

[0076] The structures of the first backbone network, the second backbone network and the third backbone network are the same;

[0077] The first backbone network includes an input layer, a first bottleneck layer, a second bottleneck layer, a third bottleneck layer, a fourth bottleneck layer and an output layer which are connected in sequence.

[0078] Specifically, the first identification head and the second identification head have the same structure. Figure 3 As shown in (a), the receptive field module is Figure 3 (b) is shown. The structure of the backbone network is shown in Figure 4 As shown in the figure, the backbone network uses the first 5 levels of the ResNet-50 network, where the second, third, fourth, and fifth levels have 3, 4, 6, and 3 bottleneck layers, respectively. In order to reduce the number of parameters, the three-step convolution of the bottleneck layer first reduces the dimension of the data by point convolution, then performs convolution with a conventional convolution kernel, and finally increases the dimension of the data by point convolution, using multiple small convolution kernels to replace a large convolution kernel, and using a 1x1 convolution kernel to replace part of the work of the large convolution kernel. In addition, this method also adds a 3×3 filter with a step size of 2 at level 1 to provide different receptive fields, which helps capture pedestrians of different scales. In the 6-level backbone network, we use the last 4 levels as detection outputs.

[0079] Furthermore, this method first uses the loss L VE Train the visible target positioning and obtain the visible part anchor point A1. Then, the key part positioning association network converts A1 into an anchor point A2 with stronger adaptability, so as to better approximate the whole body area. Furthermore, the obtained feature map is associated with key parts, each human body is divided into three hierarchical parts, and the offsets of all key points are regressed based on the center position of each part to form the initial posture. After network correction, redundant postures are eliminated, and finally the predicted posture of the human body is obtained and the key points are located. Finally, under the initialization of A2, the loss L is used. FR Training for localization of invisible parts.

[0080] Step S1 is specifically as follows:

[0081] S11: input the training image into the single-stage recognizer to obtain a first size feature map, a second size feature map, a third size feature map, and a fourth size feature map;

[0082] S12: The first recognition channel, the second recognition channel, the third recognition channel and the fourth recognition channel are trained respectively through the first size feature map, the second size feature map, the third size feature map and the fourth size feature map to obtain a trained real-time occluded target recognition model.

[0083] Furthermore, the training processes of the first recognition channel, the second recognition channel, the third recognition channel and the fourth recognition channel are the same. The training process of the first recognition channel is specifically as follows:

[0084] S121: Set a pre-anchor point A0, input the first size feature map into the first backbone network to obtain a visible part image, input the visible part image and the pre-anchor point A0 into the first recognition head, obtain a visible part anchor point A1 and a visible classification result, and calculate the loss L of the visible classification result VE ;

[0085] Furthermore, in order to train the visible target location E v (.), Our method adopts the standard recognition method, which learns to locate anchor point A1 as regression (from the predefined anchor point A0) and uses the anchor point score as classification.

[0086] Loss L VE The calculation formula is:

[0087] L VE =L local +λ v [y=1]L smoothL1

[0088] Among them, L local is the focal loss of the visible classification result, L smoothL1 is the smooth-L1 loss of A1 regression, [y=1] is the indicator of positive samples, λ v is the first tuning parameter.

[0089] Specifically, since visible object localization is trained on visible parts, its predictions on visible parts are usually more confident and accurate than recognizers trained with occlusions.

[0090] S122: Repeat step S121 until the loss is L VE Less than the preset value, a trained visible target positioning network is obtained;

[0091] S123: Divide the human body into three scales, including: key point scale s1, low-level local scale s2 and high-level local scale s3, and calculate the semantic local feature V of the key point scale s1 s1 and the global feature V g ;

[0092] Furthermore, this method uses a convolutional network to predict the human key points in the above generated feature map, and the backbone network takes the image or video frame as input to extract the initial feature map Where H, W and C represent the height, width and channel size of the feature map respectively. A key point recognition network is used to adaptively locate the key points of the human body. The key point recognition network can adaptively locate the key points of the human body and extract aligned local features from these key points to cope with the problems of partial occlusion, misalignment and viewpoint changes.

[0093] The method starts with a high-resolution subnet as the initial stage, and gradually adds subnets from high resolution to low resolution, building a multi-stage, parallel subnet structure with multiple resolutions. The resolution of subsequent parallel subnets depends on the resolution of the previous stage and the newly added low-resolution subnet. The following shows an example network structure with 4 parallel subnets:

[0094] N 11 →N 21 →N 31 →N 41

[0095] N 22 →N 32 →N 42

[0096] N 33 →N 43

[0097] N 44

[0098] Exchange units are introduced in these parallel subnetworks, so that each subnetwork can continuously receive information exchanges from other parallel subnetworks. Below is an example showing the scheme of exchanging information. The method subdivides this stage into multiple exchange blocks, each of which consists of multiple parallel convolutional units, and exchange units are set between these parallel units. The following is an example of 3 parallel units:

[0099]

[0100] in represents the convolutional unit of the resolution of level r of the b-th block in level s, and is the corresponding exchange unit.

[0101] As shown in the above exchange unit, the input is s response graphs: {X1,X2,...X s}, the output is s corresponding graphs: {Y1,Y2,...Y s}, whose resolution and width are the same as the input, and each output is an aggregation of the input:

[0102]

[0103] The cross-stage exchange unit also has an additional output map Y s+1

[0104] Y s+1 =a(Y s ,s+1)

[0105] Function a(X i ,k) contains pairs of X from i to k iFor upsampling and downsampling, 3×3 piecewise convolution is used for downsampling. For example, a 3×3 convolution with a stride of 2 is used for 2x downsampling, while two consecutive 3×3 convolutions with a stride of 2 can achieve 4x downsampling. For upsampling, 1×1 convolution is used in combination with simple nearest neighbor interpolation to ensure the alignment of the number of channels. When i=k, the function a is only connected to one identifier, a(X i ,k)=X i .

[0106] Furthermore, a heatmap regression is performed based on the high-resolution representation output by the last exchange unit. The loss function uses the mean squared error to evaluate the difference between the predicted heatmap and the true heatmap. The true heatmap is generated by a 2D Gaussian distribution with a standard deviation set to 1 pixel, centered on the true position of each keypoint.

[0107] Although keypoint estimation models have achieved high accuracy, they still suffer from unstable performance in complex surveillance scenarios, leading to inaccurate keypoint localization of occluded objects and insufficient confidence. Therefore, it is particularly important to explore multi-scale local features with spatiotemporal correlations, which helps reduce unreliable keypoint estimation results and capture different discriminative semantic information.

[0108] According to the characteristics of the human body, this method divides the human body into three granularities: key point scale (s1), low-level local scale (s2) and high-level local scale (s3). This method merges key points near the space, and each part is based on the coarse-scale human body prior. The key point estimator generates a heat map Y of the key parts of the human body, and the softmax function is used to normalize the generated heat map.

[0109] Semantic local feature V s1 and the global feature V g The calculation formula is:

[0110]

[0111] V g =g GAP (F)

[0112] in, and g GAP They refer to the outer product and global average pooling operations respectively; F is the feature map obtained after the visible part image is input into the second backbone network, It is a heat map of key parts of the human body obtained by the key part recognition network.

[0113] S124: By using the semantic local feature V s1Perform an average pooling operation to calculate the semantic local features V of the low-level local scale s2 s2 and the semantic local features V of the high-level local scale s3 s3 ;

[0114] S125: Through semantic local features V s1 、V s2 、V s3 and the global feature V g The visible part of the image is segmented into human bodies and input into the key part positioning association network to obtain the initial whole body anchor point A2;

[0115] Specifically, after obtaining a reliable visible part anchor point A1 after locating the visible target, this method uses a simple and effective key part location association method C(.) to migrate the visible part anchor point to the whole body anchor point A2, and then pass it to the next stage for further refinement. The specific process is: Since the aspect ratio of the visible part recognition box is quite different from that of the full body recognition box, the regression from the visible part to the full body becomes quite challenging. Adopting adaptive anchor initialization can reduce the unnecessary search space and achieve better recognition results. In addition, since the intersection over union (IoU) between the visible part anchor and the full body ground truth box is quite different, proper calibration can significantly improve the IoU.

[0116] This method divides each human body into three levels of components, regresses the offsets of all key points of each person based on the centers of the three levels of components, assembles them into an initial posture, and then corrects them through the network to eliminate redundant postures, obtains the final predicted human posture and locates the key points.

[0117] In order to cope with crowded scenes and accurately predict the center of the human body even when it is occluded, this method divides each human instance into three hierarchical components: head, torso, and legs. A center is defined in each hierarchical component, and its corresponding hierarchical component center heat map and offset prediction map are recorded as Where Y is the part center heat map, O is the offset prediction map, M is the number of hierarchical components, and p is the component.

[0118] When the center of a part is occluded, for example, the key point of the head is occluded, it is difficult for the network to regress the offset based on the center of the head part. However, due to the structured relationship between the key points of the human body, the offsets based on the regression of the center of the torso part and the center of the leg part have a specific vector relationship in space. At this time, the posture obtained based on the regression of the center of the multi-level components enhances the robustness of recognition in crowded scenes. By combining the center heat map Y and the offset prediction map O of the obtained hierarchical parts into a posture template and comparing and correcting it with the established posture template (supplemented by collecting labeled human data sets for different postures of the human body), after removing redundant postures, and selecting the key points with the highest scores in each type of key points from all postures in turn as the key points of the final posture in human occlusion recognition. Thereby obtaining a new predicted human posture and further performing identification marking of the key points of the human body.

[0119] S126: Input the first size feature map into the third backbone network to obtain a full-body image, input the full-body image and the initial full-body anchor point A2 into the second recognition head, obtain the final full-body anchor point A3 and the full-body classification result, and calculate the loss L of the full-body classification result FR ;

[0120] Furthermore, using the anchor point A2 associated with the calibration of the important part positioning, the final stage of recognition will train an invisible part locator E f (.) to refine the final full-body localization. Similar to visible object localization, invisible part localization also uses the same backbone network but is trained on a separate recognition head. Unlike visible object localization which only sees the visible part, invisible part localization starts to see hard positive samples where the anchor box is still far from the true ground-truth full-body region. Since LsmoothL1 performs the same treatment on every positive sample, its effect may be greatly reduced when dealing with hard samples in invisible part localization. Given a∈A2 and its corresponding b∈B gt .

[0121] Loss L FR The calculation formula is:

[0122] L FR =L local +λ f [y=1]L occ

[0123]

[0124] Among them, L local is the focal loss of the visible classification result, [y=1] is the indicator of positive samples, λ f is the second tuning parameter; L occ is the weighted loss, a is the anchor box of the initial full-body anchor A2, b∈Bgt , B gt is the set of true bounding boxes, s is the difference between the predicted offset and the true ground-truth offset, and IoU(a,b) is the intersection over union (IoU) between the anchor box of the initial full-body anchor and the true bounding box.

[0125] Specifically, the initial whole body anchor points A2 and B gt The less overlap there is, the better occ The higher the value. Visible object localization and invisible part localization use two different recognition heads in the same single-stage recognizer, so their learning outputs are complementary. When the intersection over union (IoU) between the anchor box and the true box exceeds the set threshold, the anchor is marked as a positive sample; if it is lower than the threshold, it is marked as a negative sample; otherwise, it will be ignored during training.

[0126] Denote x as the input image, φ(x) as the feature map from the backbone, A0 as a set of predefined anchor points, and B0 as the predicted bounding box obtained by non-maximum suppression of the anchor points collected from all layers. Given an initial set of feature maps and anchor points, the recognizer can be formulated as a progressive recognizer:

[0127] A3=F(E f (C(E v (φ(x),A0))))={B * ,s *}

[0128] Where E v (φ(x), A0) is the first stage visible object localization, whose output is a set of visible part anchors and confidence scores {A1, s1}, C(.) is the second stage key part localization association that aligns the visible part anchor A1 with the initial whole body anchor A2, and E f (φ(x),A2) is the third stage invisible part localizer, which outputs the final full body anchor to calculate B* and its score s* using reasoning F, φ(x) represents the different feature maps during visible object localization and invisible part localization.

[0129] S127: Repeat step S126 until the loss is L FR is less than a preset value, a trained invisible part positioning network is obtained, and a trained first recognition channel is output.

[0130] S2: Acquire the image of the area to be identified and perform preprocessing to generate a preprocessed image;

[0131] Specifically, a high-resolution camera is used to capture images of the occluded area in real time and preprocess them. The camera has a wide-angle field of view and night vision, and can capture images clearly under different lighting conditions. The original image undergoes a series of preprocessing steps, including image scaling, normalization, denoising, etc., to adjust the image size and pixel value range to meet the input requirements of the subsequent deep learning model. Preprocessing also includes operations such as enhancing the contrast and sharpness of the image and flipping and shearing the image to improve image quality and feature extraction effects.

[0132] S3: Input the preprocessed image into the trained real-time occluded object recognition model to obtain the comprehensive final full-body anchor point and comprehensive full-body classification results;

[0133] Furthermore, step S3 is specifically as follows:

[0134] S31: Input the preprocessed image into the trained real-time occluded target recognition model, and obtain the first final whole-body anchor point, the first whole-body classification result, the second final whole-body anchor point, the second whole-body classification result, the third final whole-body anchor point, the third whole-body classification result, the fourth final whole-body anchor point, and the fourth whole-body classification result through the first recognition channel, the second recognition channel, the third recognition channel, and the fourth recognition channel respectively;

[0135] S32: adjusting the first final whole-body anchor point and the first whole-body classification result through the first receptive field module, adjusting the second final whole-body anchor point and the second whole-body classification result through the second receptive field module, and adjusting the third final whole-body anchor point and the third whole-body classification result through the third receptive field module;

[0136] Specifically, despite increasing the weights of hard positive anchors, another challenge of invisible part localization lies in the training of shallow layers. Since the receptive field of shallow networks is limited, they are usually activated in visible parts or small-sized full-body regions. To address this problem, a receptive field module is introduced in each layer of invisible part localization to diversify the receptive field. The receptive field module expands the receptive field of the shallow layer by upsampling the features of the deeper layer by a factor of 2 and feeding it back to the previous layer, followed by pixel-level aggregation of the feature map.

[0137] S33: Fuse the first final whole-body anchor point, the second final whole-body anchor point, the third final whole-body anchor point and the fourth final whole-body anchor point to obtain a comprehensive final whole-body anchor point; fuse the first whole-body classification result, the second whole-body classification result, the third whole-body classification result and the fourth whole-body classification result to obtain a comprehensive whole-body classification result.

[0138] S4: Analyze the comprehensive final whole-body anchor points and the comprehensive whole-body classification results to obtain the key parts of the target.

[0139] Specifically, further data analysis can be performed on the key parts of the target to obtain the target's location and category data, risk data, and warning information, and these information can be displayed visually.

[0140] See also Figure 5 , Figure 5 4 is a schematic diagram of the working of the hardware device of an embodiment of the present invention, and the hardware device specifically includes: an intelligent recognition device 401 for key parts of a hidden target, a processor 402 and a storage medium 403.

[0141] An intelligent device 401 for identifying key parts of a hidden target: The intelligent device 401 for identifying key parts of a hidden target implements the intelligent method for identifying key parts of a hidden target.

[0142] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the method for intelligently identifying the key parts of the concealed target.

[0143] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the intelligent identification method of the key parts of the hidden target.

[0144] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0145] The serial numbers of the embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In a unit claim that lists several means, several of these means may be embodied by the same hardware item. The use of the words first, second, and third, etc. does not indicate any order and these words may be interpreted as identifiers.

[0146] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for intelligently identifying key parts of a hidden target, characterized in that: The following steps are involved: S1: construct a real-time occluded target recognition model, train the real-time occluded target recognition model through training images, and obtain a trained real-time occluded target recognition model; S2: Acquire the image of the area to be identified and perform preprocessing to generate a preprocessed image; S3: Input the preprocessed image into the trained real-time occluded object recognition model to obtain the comprehensive final full-body anchor point and comprehensive full-body classification results; S4: Analyze the comprehensive final whole-body anchor points and the comprehensive whole-body classification results to obtain the key parts of the target.

2. The method for intelligently identifying key parts of a concealed target according to claim 1, characterized in that: The real-time occluded target recognition model includes: a single-stage recognizer, a first recognition channel, a second recognition channel, a third recognition channel, a fourth recognition channel, a first receptive field module, a second receptive field module and a third receptive field module; The single-stage identifier is connected to the first identification channel, the second identification channel, the third identification channel and the fourth identification channel; The output end of the fourth identification channel, the third receptive field module, the output end of the third identification channel, the second receptive field module, the output end of the second identification channel, the first receptive field module and the output end of the first identification channel are connected in sequence.

3. The method for intelligently identifying key parts of a concealed target according to claim 2, characterized in that: The first identification channel, the second identification channel, the third identification channel and the fourth identification channel have the same structure; The first recognition channel includes a visible target positioning network, a key part positioning association network and an invisible part positioning network connected in sequence; The visible target positioning network includes: a first backbone network and a first recognition head; The key part positioning association network includes: a second backbone network and a key part identification network; The invisible part localization network includes: a third backbone network and a second recognition head; The structures of the first backbone network, the second backbone network and the third backbone network are the same; The first backbone network includes an input layer, a first bottleneck layer, a second bottleneck layer, a third bottleneck layer, a fourth bottleneck layer and an output layer which are connected in sequence.

4. The method for intelligently identifying key parts of a hidden target according to claim 3, characterized in that: Step S1 is specifically as follows: S11: input the training image into the single-stage recognizer to obtain a first size feature map, a second size feature map, a third size feature map, and a fourth size feature map; S12: The first recognition channel, the second recognition channel, the third recognition channel and the fourth recognition channel are trained respectively through the first size feature map, the second size feature map, the third size feature map and the fourth size feature map to obtain a trained real-time occluded target recognition model.

5. The method for intelligently identifying key parts of a hidden target according to claim 4, characterized in that: The training process of the first recognition channel, the second recognition channel, the third recognition channel and the fourth recognition channel is the same. The training process of the first recognition channel is specifically as follows: S121: Set a pre-anchor point A0, input the first size feature map into the first backbone network to obtain a visible part image, input the visible part image and the pre-anchor point A0 into the first recognition head, obtain a visible part anchor point A1 and a visible classification result, and calculate the loss L of the visible classification result VE ; S122: Repeat step S121 until the loss is L VE Less than the preset value, a trained visible target positioning network is obtained; S123: Divide the human body into three scales, including: key point scale s1, low-level local scale s2 and high-level local scale s3, and calculate the semantic local feature V of the key point scale s1 s1 and the global feature V g ; S124: By using the semantic local feature V s1 Perform an average pooling operation to calculate the semantic local features V of the low-level local scale s2 s2 and the semantic local features V of the high-level local scale s3 s3 ; S125: Through semantic local features V s1 、V s2 、V s3 and the global feature V g The visible part of the image is segmented into human bodies and input into the key part positioning association network to obtain the initial whole body anchor point A2; S126: Input the first size feature map into the third backbone network to obtain a full-body image, input the full-body image and the initial full-body anchor point A2 into the second recognition head, obtain the final full-body anchor point A3 and the full-body classification result, and calculate the loss L of the full-body classification result FR ; S127: Repeat step S126 until the loss is L FR is less than a preset value, a trained invisible part positioning network is obtained, and a trained first recognition channel is output.

6. The method for intelligently identifying key parts of a concealed target according to claim 5, characterized in that: Loss L VE The calculation formula is: L VE =L local +λ v [y=1]L smoothL1 Among them, L local is the focal loss of the visible classification result, L smoothL1 is the smooth-L1 loss of A1 regression, [y=1] is the indicator of positive samples, λ v is the first tuning parameter.

7. The method for intelligently identifying key parts of a concealed target according to claim 5, characterized in that: Semantic local feature V s1 and the global feature V g The calculation formula is: V g =g GAP (F) in, and g GAP They refer to the outer product and global average pooling operations respectively; F is the feature map obtained after the visible part image is input into the second backbone network, It is a heat map of key parts of the human body obtained by the key part recognition network.

8. The method for intelligently identifying key parts of a concealed target according to claim 5, characterized in that: Loss L FR The calculation formula is: L FR =L local +λ f [y=1]L occ Among them, L local is the focal loss of the visible classification result, [y=1] is the indicator of positive samples, λ f is the second tuning parameter; L occ is the weighted loss, a is the anchor box of the initial full-body anchor A2, b∈B gt , B gt is the set of true bounding boxes, s is the difference between the predicted offset and the true ground-truth offset, and IoU(a,b) is the intersection over union (IoU) between the anchor box of the initial full-body anchor and the true bounding box.

9. The method for intelligently identifying key parts of a hidden target according to claim 3, characterized in that: Step S3 is specifically as follows: S31: Input the preprocessed image into the trained real-time occluded target recognition model, and obtain the first final whole-body anchor point, the first whole-body classification result, the second final whole-body anchor point, the second whole-body classification result, the third final whole-body anchor point, the third whole-body classification result, the fourth final whole-body anchor point, and the fourth whole-body classification result through the first recognition channel, the second recognition channel, the third recognition channel, and the fourth recognition channel respectively; S32: adjusting the first final whole-body anchor point and the first whole-body classification result through the first receptive field module, adjusting the second final whole-body anchor point and the second whole-body classification result through the second receptive field module, and adjusting the third final whole-body anchor point and the third whole-body classification result through the third receptive field module; S33: fusing the first final whole body anchor point, the second final whole body anchor point, the third final whole body anchor point and the fourth final whole body anchor point to obtain a comprehensive final whole body anchor point; The first whole-body classification result, the second whole-body classification result, the third whole-body classification result and the fourth whole-body classification result are integrated to obtain a comprehensive whole-body classification result.

10. An intelligent identification device for key parts of a concealed target, characterized in that: include: Processor and storage medium; the processor loads and executes instructions and data in the storage medium to implement the method for intelligently identifying key parts of a concealed target as described in any one of claims 1 to 9.

Images