A human recognition method, device and equipment for top view angle and medium

Abstract

The application discloses a human body recognition method and device for top view angle, equipment and medium, including: obtaining a target image; based on the frame prediction head and the point prediction head in the preset recognition model, the target image is identified, the human body detection frame and the head vertex coordinates of the multiple scale feature maps in the preset recognition model are obtained; based on the human body detection frame of multiple scale feature maps, the small scale feature map in the multiple scale feature map is processed, and the target small scale feature map is obtained; based on the appearance representation prediction head in the preset recognition model, the target small scale feature map is processed, and the target representation vector is obtained, which is used for recognizing the human body. The human body detection frame and the head vertex coordinates are used for positioning, which can improve the recall rate and accuracy of pedestrians, and the output appearance representation vector is used for human frame matching, which saves the time-consuming of separately extracting the appearance representation vector and improves the efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, and in particular to a human body recognition method and device for top view, equipment and medium. BACKGROUND

[0002] With the development of deep learning, more and more fields are using target recognition technology, and the target recognition of human body can be used in security inspection, identity verification, mobile payment and other fields. Multi-target tracking refers to identifying and tracking multiple objects without knowing the appearance and number of the targets.

[0003] In order to improve the safety of people's daily activities, intelligent monitoring has been widely used in the security field to provide protection for people's property and life safety. The monitoring camera is usually installed at a downward or top view angle. When detecting and tracking pedestrians, the body is usually recognized, which can easily lead to background false positives, the generalization ability of the recognition model is weak, and the accuracy is not high. SUMMARY

[0004] Therefore, the embodiments of the present application provide a human body recognition method and device for top view, equipment and medium to solve the problem of low efficiency in positioning the human body during top view recognition of pedestrians.

[0005] According to a first aspect, the embodiments of the present application provide a human body recognition method for top view, comprising:

[0006] obtaining a target image;

[0007] recognizing the target image based on a box prediction head and a point prediction head in a preset recognition model to obtain a human body detection box and a head point coordinate of a plurality of scale feature maps in the preset recognition model;

[0008] processing a small scale feature map in the plurality of scale feature maps based on the human body detection box of the plurality of scale feature maps to obtain a target small scale feature map;

[0009] processing the target small scale feature map based on an appearance representation prediction head in the preset recognition model to obtain a target representation vector for recognizing the human body.

[0010] The human body recognition method for top view provided by the embodiments of the present application recognizes the target image, the box prediction head and the point prediction head output the human body detection box and the head point coordinate, and the appearance representation prediction head outputs the appearance representation vector. The human body detection box and the head point coordinate are used for positioning, which can improve the recall rate and accuracy of pedestrians, and the output appearance representation vector is used for human frame matching, which saves the time-consuming of separately extracting the appearance representation vector and improves the efficiency.

[0011] In some embodiments, the plurality of scale feature maps comprises a large scale feature map, a medium scale feature map and a small scale feature map; the human body bounding box based on the plurality of scale feature maps processes the small scale feature map in the plurality of scale feature maps to obtain a target small scale feature map, comprising:

[0012] mapping the human body bounding box in the large scale feature map and the medium scale feature map to the small scale feature map to obtain the target small scale feature map.

[0013] In some embodiments, the large scale feature map, the medium scale feature map and the small scale feature map respectively comprise a box prediction head and a point prediction head; the target image is identified based on the box prediction head and the point prediction head in the preset identification model to obtain the human body bounding box and the head vertex coordinates of the plurality of scale feature maps in the preset identification model, comprising:

[0014] processing the large scale feature image, the medium scale feature image and the small scale feature image based on each of the box prediction head and the point prediction head to obtain the human body bounding box, the head vertex coordinates, the bounding box confidence and the head vertex confidence of the plurality of scale feature maps.

[0015] In some embodiments, the target small scale feature map is processed based on the appearance representation prediction head in the preset identification model to obtain a target representation vector, comprising:

[0016] cropping the target small scale feature map to obtain a target region image;

[0017] identifying the target region image based on the appearance representation prediction head in the preset identification model to obtain the target representation vector.

[0018] In some embodiments, the preset identification model is trained according to the following method:

[0019] obtaining a labeled image, the labeled image comprising a bounding box, a head vertex and an identification number of each target object;

[0020] processing the labeled image based on each of the box prediction head and the point prediction head, the box prediction head processing to obtain a box coordinate and a box confidence, and the point prediction head processing to obtain a head vertex coordinate and a head vertex confidence;

[0021] processing the small scale feature image based on each box coordinate to obtain a target feature image;

[0022] cropping the target feature image to obtain a target region, and inputting the target region into the appearance representation prediction head to obtain an appearance representation vector;

[0023] construct a target loss function based on the bounding box coordinates, the bounding box confidence, the head vertex coordinates, the head vertex confidence, and the appearance feature vector, and optimize the target loss function to determine the preset recognition model.

[0024] In some embodiments, the bounding box confidence includes a bounding box category confidence and a bounding box target confidence, and the constructing the target loss function based on the bounding box coordinates, the bounding box confidence, the head vertex coordinates, the head vertex confidence, and the appearance feature vector includes:

[0025] respectively, to obtain a bounding box coordinate loss, a bounding box target confidence loss, a bounding box category confidence loss, a head vertex coordinate loss, a head vertex confidence loss, and an appearance feature vector loss;

[0026] The target loss function is determined according to the following formula:

[0027]

[0028] wherein, L total represents the target loss function, M represents the number of scales, respectively, represent a bounding box coordinate loss, a bounding box target confidence loss, a bounding box category confidence loss, a head vertex confidence loss, a head vertex coordinate loss, and an appearance feature vector loss, represents a weighting parameter related to each loss.

[0029] In some embodiments, the preset recognition model includes a large-scale feature image, a medium-scale feature image, and a small-scale feature image, and the processing the small-scale feature image based on each bounding box coordinate and each head vertex coordinate to obtain a target feature image includes:

[0030] respectively, to obtain a target feature image.

[0031] According to a second aspect, an embodiment of the present application provides a human body recognition device for top-view angle, comprising:

[0032] an image acquisition module, configured to acquire a target image;

[0033] an image recognition module, configured to recognize the target image based on a bounding box prediction head and a point prediction head in a preset recognition model to obtain a human body detection bounding box and a head vertex coordinate of a plurality of scale feature images in the preset recognition model;

[0034] The feature processing module is configured to process a small-scale feature map in the plurality of scale feature maps based on a human body bounding box of the plurality of scale feature maps, to obtain a target small-scale feature map.

[0035] The appearance representation processing module is configured to process the target small-scale feature map based on an appearance representation prediction head in the preset identification model, to obtain a target representation vector, for identifying a human body.

[0036] According to a third aspect, an embodiment of the present application provides an electronic device, comprising a memory and a processor, which are in communication connection with each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the human body identification method for top view angle according to the first aspect or any one of the implementation manners of the first aspect.

[0037] According to a fourth aspect, an embodiment of the present application provides a computer readable storage medium, which stores computer instructions for causing the computer to perform the human body identification method for top view angle according to the first aspect or any one of the implementation manners of the first aspect. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the drawings needed in the specific embodiments or the prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0039] Figure 1 is a flowchart of the human body identification method for top view angle according to an embodiment of the present application;

[0040] Figure 2 is a schematic diagram of a target image according to an embodiment of the present application;

[0041] Figure 3 is a model framework diagram according to an embodiment of the present application;

[0042] Figure 4 is an effect schematic diagram of the human body identification method for top view angle according to an embodiment of the present application;

[0043] Figure 5 is a schematic diagram of the human body identification device for top view angle according to an embodiment of the present application;

[0044] Figure 6 is a hardware structure schematic diagram of the electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0045] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0046] According to the embodiments of the present application, an embodiment of a human body recognition method for top view is provided. It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in a different order.

[0047] In the present embodiment, a human body recognition method for top view is provided, Figure 1 is a flowchart of the human body recognition method for top view according to the embodiments of the present application, as Figure 1 shown, the flow includes the following steps:

[0048] S11, obtaining a target image.

[0049] Collecting image data of top view scenes, overhead view scenes, or video data similar to top view and overhead view scenes to obtain a target image. The target image is as Figure 2 shown.

[0050] S12, identifying the target image based on a box prediction head and a point prediction head in a preset recognition model to obtain a human body detection box and a head point coordinate of a plurality of scale feature maps in the preset recognition model.

[0051] The preset recognition model is a pre-trained target recognition model. The Joint Detection and Embedding (JDE) model is added with a box prediction head and a point prediction head. The box prediction head is used to predict the human body detection box, and the point prediction head is used to predict the head point coordinate of the pedestrian in the image. The target image is input into the preset recognition model to first obtain a plurality of scale feature maps. The number of scales is not limited herein. The human body detection box and the head point coordinate of each scale feature map are output. The output human body detection box coordinate is relative to the center point coordinate, width and height of the preset anchor. The output effect diagram is as Figure 4 shown.

[0052] S13, processing a small scale feature map in the plurality of scale feature maps based on the human body detection box of the plurality of scale feature maps to obtain a target small scale feature map.

[0053] Map the human body detection frame output by each scale feature map to the feature map with the largest resolution and the smallest scale to obtain a target small-scale feature map.

[0054] S14, processing the target small-scale feature map based on an appearance feature prediction head in a preset identification model to obtain a target feature vector for identifying the human body.

[0055] The appearance feature prediction head in the preset identification model can first crop the target small-scale feature map, and the ROI Align algorithm can be used during the cropping to obtain the human body detection frame region in the target small-scale feature map. The cropped target small-scale feature map is input into the appearance feature prediction head to output the target feature vector, which is an appearance feature vector. In actual use, the similarity of the appearance feature vectors of the human body frames can be compared to distinguish whether they are the same person, which is used for subsequent positioning of the human body in the target image.

[0056] The human body identification method for top view provided in the embodiment is used for identifying a target image. The human body detection frame and the head vertex coordinate output by each frame prediction head and point prediction head are used for positioning, which can improve the recall rate and accuracy of pedestrians. The appearance feature vector output at the same time is used for frame matching, which saves the time-consuming of separately extracting the appearance feature vector and improves the efficiency.

[0057] In some embodiments, the plurality of scale feature maps include a large-scale feature map, a medium-scale feature map, and a small-scale feature map. Figure 1 S13 includes: mapping the human body detection frame and the head vertex coordinate in the large-scale feature map and the medium-scale feature map to the small-scale feature map to obtain a target small-scale feature map.

[0058] In this embodiment, taking the plurality of scale feature maps including the large-scale feature map, the medium-scale feature map, and the small-scale feature map as an example, it is assumed that the small-scale feature map is one-eighth of the network input, the medium-scale feature map is one-sixteenth of the network input, and the coordinate x2 predicted on the medium-scale feature map corresponds to the coordinate on the small-scale feature map.

[0059] In some embodiments, the large-scale feature map, the medium-scale feature map, and the small-scale feature map each include a frame prediction head and a point prediction head. Figure 1 S12 includes:

[0060] Processing the large-scale feature image, the medium-scale feature image, and the small-scale feature image based on each frame prediction head and point prediction head to obtain the human body detection frame, the head vertex coordinate, the detection frame confidence, and the head vertex confidence of the plurality of scale feature maps.

[0061] As shown in the model framework diagram, Figure 3 the corresponding frame prediction head and point prediction head are connected in the large-scale feature map, the medium-scale feature map and the small-scale feature map, and the feature maps of each scale are predicted respectively. Each grid in the frame prediction head predicts three anchors, each anchor predicts a human detection frame and a detection frame confidence. The human detection frame is a four-dimensional frame coordinate, including the center point coordinate, the width and the height relative to the anchor. The detection frame confidence can include a one-dimensional class confidence and a one-dimensional target confidence. The target confidence represents the confidence that a certain predicted frame has a target, and the class confidence represents the confidence that a certain class belongs to when there is a target. The final detection frame confidence is equal to the product of the class confidence and the target confidence. In this embodiment, the class confidence is 1, indicating that there is only one class of pedestrians. Each grid in the point prediction head corresponds to the grid in the frame prediction head, and predicts the head vertex coordinates and the head vertex confidence of the corresponding anchor. The head vertex coordinates are two-dimensional, and the head vertex confidence is one-dimensional.

[0062] In some embodiments, Figure 1 S14 in the method comprises the following steps:

[0063] S21, cropping a target small-scale feature map to obtain a target region image.

[0064] S22, identifying the target region image based on the appearance feature prediction head in the preset identification model to obtain a target feature vector.

[0065] The embedding branch of the JDE model is changed to an ROI Align branch, and the target small-scale feature map is processed based on the ROI Align algorithm. The corresponding region of the human detection frame in the target small-scale feature map, i.e. the target region image, is cropped. The target region image is input into the appearance feature prediction head in the model to obtain the target feature vector.

[0066] In some embodiments, Figure 1 The preset identification model in the method is trained according to the following method:

[0067] S31, obtaining a labeled image, the labeled image including a detection frame, a head vertex and an identification number of each target object.

[0068] Collect image data or video data of a top view or a bird's eye view of a target scene, sample and label the video data, and label the image data. Each image is labeled with a detection frame, a head vertex position, and an identification number of each target object to produce a training data set and a test data set. The target object in this embodiment is a pedestrian, and the detection frame labels the upper left corner coordinates and the lower right corner coordinates of the human detection frame.

[0069] S32, the labeled image is processed based on each bounding box prediction head and point prediction head. The bounding box prediction head is processed to obtain the bounding box coordinates and bounding box confidence, and the point prediction head is processed to obtain the head vertex coordinates and head vertex confidence.

[0070] A training model is constructed, which includes a backbone network. After the labeled image is input into the model, it passes through the backbone network and outputs feature images at multiple scales. Each feature image at a scale is connected to a bounding box prediction head and a point prediction head. The bounding box prediction head outputs the bounding box coordinates and the bounding box confidence score, while the point prediction head outputs the point vertex coordinates and the point vertex confidence score.

[0071] by Figure 3 Taking the model network structure shown as an example, the embedding branch of JDE is changed to the ROI Align branch. After the image is input into the model, it passes through the backbone network and outputs large-scale feature images, medium-scale feature images, and small-scale feature images. Each scale feature image is connected to the bounding box prediction head and the point prediction head. Prediction is performed on the feature maps of each scale. Each grid in the bounding box prediction head predicts three anchors. Each anchor predicts four-dimensional bounding box coordinates and two-dimensional bounding box confidence. The bounding box coordinates include the coordinates of the center point relative to the anchor, the width, and the height. The bounding box confidence can include one-dimensional class confidence and one-dimensional target confidence. The target confidence represents the confidence that a predicted bounding box has a target, and the class confidence represents the confidence that it belongs to a certain class when there is a target. The final bounding box confidence is equal to the product of the class confidence and the target confidence. In this embodiment, the class confidence is 1, indicating that there is only one category: pedestrian. Each grid in the point prediction head corresponds to a grid in the box prediction head, predicting the head vertex coordinates and head vertex confidence of the corresponding anchor. The head vertex coordinates are two-dimensional, and the head vertex confidence is one-dimensional.

[0072] S33. Based on the coordinates of each bounding box and the coordinates of each head vertex, the small-scale feature image is processed to obtain the target feature image.

[0073] In some implementations, the preset recognition model includes large-scale feature images, medium-scale feature images, and small-scale feature images, and S33 includes:

[0074] The bounding box coordinates in the large-scale feature image and the bounding box coordinates and head vertex coordinates in the medium-scale feature image are respectively mapped to the small-scale feature image to obtain the target feature image. Assuming that the small-scale feature map is one-eighth of the network input and the medium-scale feature map is one-sixteenth of the network input, the predicted coordinates x2 on the medium-scale feature map will yield the corresponding coordinates on the small-scale feature map.

[0075] S34, crop the target feature image to obtain the target region, and input the target region into the appearance representation prediction head to obtain the appearance representation vector.

[0076] The target feature image can be cropped, and the ROI Align algorithm can be used during cropping to obtain a target region in a target small-scale feature map. The target region is input into the appearance representation prediction head to output an appearance representation vector. During training, after the appearance representation prediction head obtains the appearance representation vector, the appearance representation vector needs to be input into a fully connected layer to be converted into an id prediction problem. The appearance representation vector output by the appearance representation prediction head is output through a fully connected layer to obtain a vector with a total id number. The vector and the corresponding target value are input into a CrossEntroyLoss layer to obtain the corresponding id loss. The CrossEntroyLoss layer internally performs a softmax process to convert the vector with a total id number output by the fully connected layer into a vector with a sum of 1. It can be understood that at this time, each value of the vector represents the confidence of the id, and then the id loss is calculated.

[0077] S35, based on the box coordinates, the box confidence, the head vertex coordinates, the head vertex confidence and the appearance representation vector, a target loss function is constructed, and the target loss function is optimized to determine the preset identification model.

[0078] The obtained box coordinates and head vertex coordinates are coordinate data relative to the center point of the anchor. In training, the anchor is assigned an id to avoid the problem that one grid corresponds to multiple anchors with targets and cannot be assigned an id, thereby improving the accuracy of the appearance representation vector.

[0079] The losses of the box coordinates, the box confidence, the head vertex coordinates, the head vertex confidence and the appearance representation vector are calculated respectively, and a target loss function is constructed. The target loss function is optimized.

[0080] After the model training is completed, the fully connected layer after the appearance representation prediction head is removed, and the appearance representation vector is directly taken, or the fully connected layer is defined at the place where the loss is calculated, and the network directly outputs the appearance representation vector. In this embodiment, the confidence threshold of the head vertex confidence is set to 0.5. When the head vertex confidence is greater than 0.5, the head vertex is considered to be reliable.

[0081] In some embodiments, the box confidence includes a box category confidence and a box target confidence, and S35 includes: calculating the losses based on the box coordinates, the box confidence, the head vertex coordinates, the head vertex confidence and the appearance representation vector respectively to obtain a box coordinate loss, a box target confidence loss, a box category confidence loss, a head vertex coordinate loss, a head vertex confidence loss and an appearance representation vector loss.

[0082] The target loss function is determined according to the following formula:

[0083]

[0084] Ltarget=λboxLbox+λheadLhead+λappearanceLappearance+λboxclassLboxclasstotal represents a target loss function, M represents the number of scales, respectively represent frame coordinate loss, frame target confidence loss, frame class confidence loss, head vertex confidence loss, head vertex coordinate loss, appearance feature vector loss, represents a weighting parameter related to each loss.

[0085] In the embodiment, the frame coordinates can be supervised by the CIOU loss, the frame class confidence and the frame target confidence adopt the BCE loss, the head vertex confidence adopts the BCE loss, the head vertex coordinates adopt the OKS loss, and the appearance feature vector adopts the CE loss.

[0086] In the human body recognition method for top view angle provided by the embodiment, the recognition model includes a frame prediction head and a point prediction head, can detect true positive cases in which the head cannot be seen but most of the body can be seen, improves the tracking recall rate, enhances the generalization ability of the model, improves the model precision, has low labeling cost and difficulty, and supports different heights. When training, the anchor is assigned an id, which avoids the problem that multiple anchors corresponding to a grid have targets and cannot be assigned an id, and improves the accuracy of the feature vector for pedestrian re-identification.

[0087] In the embodiment, a human body recognition device for top view angle is also provided, which is used to implement the above-mentioned embodiments and preferred embodiments, and will not be described again. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware, or a combination of software and hardware is also possible and is contemplated.

[0088] The embodiment provides a human body recognition device for top view angle, as shown in the figure, comprising: Figure 5

[0089] An image acquisition module is configured to acquire a target image.

[0090] An image recognition module is configured to recognize the target image based on a frame prediction head and a point prediction head in a preset recognition model, to obtain a human body detection frame and a head vertex coordinate of a plurality of scale feature maps in the preset recognition model.

[0091] A feature processing module is configured to process a small scale feature map in the plurality of scale feature maps based on the human body detection frame of the plurality of scale feature maps, to obtain a target small scale feature map.

[0092] A feature processing module is configured to process a small scale feature map in the plurality of scale feature maps based on the human body detection frame of the plurality of scale feature maps, to obtain a target small scale feature map. ​

[0093] In some embodiments, the plurality of scale feature maps include a large scale feature map, a medium scale feature map, and a small scale feature map; and the feature processing module includes:

[0094] a coordinate mapping unit, configured to map the human detection frame in the large scale feature map and the medium scale feature map to the small scale feature map to obtain a target small scale feature map.

[0095] In some embodiments, the large scale feature map, the medium scale feature map, and the small scale feature map each include a frame prediction head and a point prediction head, and the image recognition module includes:

[0096] a prediction processing unit, configured to process the large scale feature image, the medium scale feature image, and the small scale feature image based on the frame prediction head and the point prediction head respectively to obtain the human detection frame, the head vertex coordinate, the detection frame confidence, and the head vertex confidence of the plurality of scale feature maps.

[0097] In some embodiments, the feature processing module includes:

[0098] an image cropping unit, configured to crop the target small scale feature map to obtain a target region image;

[0099] an image recognition unit, configured to recognize the target region image based on an appearance feature prediction head in the preset recognition model to obtain a target feature vector.

[0100] In some embodiments, the method further includes:

[0101] a labeled image acquisition module, configured to acquire a labeled image, the labeled image including a detection frame, a head vertex, and an identification number of each target object;

[0102] a labeled image processing module, configured to process the labeled image based on the frame prediction head and the point prediction head, the frame prediction head processing to obtain frame coordinates and frame confidence, and the point prediction head processing to obtain head vertex coordinates and head vertex confidence;

[0103] a target feature image acquisition module, configured to process a small scale feature image based on the frame coordinates to obtain a target feature image;

[0104] a region cropping module, configured to crop the target feature image to obtain a target region, and input the target region into the appearance feature prediction head to obtain an appearance feature vector;

[0105] a model determination module, configured to construct a target loss function based on the frame coordinates, the frame confidence, the head vertex coordinates, the head vertex confidence, and the appearance feature vector, and optimize the target loss function to determine the preset recognition model.

[0106] In some embodiments, the box confidence includes a box category confidence and a box target confidence, and the model determination module includes:

[0107] a loss calculation unit configured to calculate the losses based on the box coordinates, the box confidence, the head vertex coordinates, the head vertex confidence, and the appearance representation vector, respectively, to obtain a box coordinate loss, a box target confidence loss, a box category confidence loss, a head vertex coordinate loss, a head vertex confidence loss, and an appearance representation vector loss;

[0108] The target loss function is determined according to the following formula:

[0109]

[0110] wherein L total represents the target loss function, M represents the number of scales, respectively represent the box coordinate loss, the box target confidence loss, the box category confidence loss, the head vertex confidence loss, the head vertex coordinate loss, and the appearance representation vector loss, represents a weighting parameter related to each loss.

[0111] In some embodiments, the preset recognition model includes a large-scale feature image, a medium-scale feature image, and a small-scale feature image, and the target feature image acquisition module includes:

[0112] a target coordinate mapping unit configured to respectively map the box coordinates in the large-scale feature image and the box coordinates and the head vertex coordinates in the medium-scale feature image to the small-scale feature image to obtain the target feature image.

[0113] The human body recognition device for top-view angle in the embodiment is presented in the form of functional units, and the units herein refer to ASIC circuits, processors and memories executing one or more software or fixed programs, and / or other devices that can provide the above functions.

[0114] Further function descriptions of the above modules are the same as those of the corresponding embodiments, and will not be described here.

[0115] The embodiment of the present application also provides an electronic device having the above Figure 5 human body recognition device for top-view angle.

[0116] Please refer to Figure 6 , Figure 6 is a structural schematic diagram of an electronic device provided by an optional embodiment of the present application, as shown in Figure 6As shown, the electronic device can include at least one processor 601, such as a CPU (Central Processing Unit), at least one communication interface 603, a memory 604, and at least one communication bus 602. The communication bus 602 is used to realize the connection and communication between the components. The communication interface 603 can include a display, a keyboard, and can also include a standard wired interface and a wireless interface. The memory 604 can be a high-speed RAM memory, and can also be a non-volatile memory, such as at least one disk memory. The memory 604 can also be at least one storage device located away from the aforementioned processor 601. The processor 601 can be combined with Figure 5 The described device, the memory 604 stores an application program, and the processor 601 invokes the program code stored in the memory 604 to execute any of the above method steps.

[0117] The communication bus 602 can be a PCI bus or an EISA bus, etc. The communication bus 602 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 6 Only one thick line is used in the figure, but it does not mean that there is only one bus or only one type of bus.

[0118] The memory 604 can include volatile memory, such as random access memory (RAM); the memory can also include non-volatile memory, such as flash memory, a hard disk (HDD) or a solid state disk (SSD); the memory 604 can also include a combination of the above types of memory.

[0119] The processor 601 can be a central processing unit (CPU), a network processor (NP), or a combination of CPU and NP.

[0120] The processor 601 can further include a hardware chip. The hardware chip can be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The PLD can be a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), a generic array logic (GAL), or any combination thereof.

[0121] Optionally, the memory 604 is further configured to store program instructions. The processor 601 can invoke the program instructions to implement the method for human body recognition from a top view as shown in the embodiments of the present application.

[0122] The embodiments of the present application further provide a non-transitory computer storage medium storing computer executable instructions. The computer executable instructions can implement the method for human body recognition from a top view in any of the method embodiments described above. The storage medium can be a disk, an optical disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a Flash Memory, a Hard Disk Drive (HDD), a Solid-State Drive (SSD), or the like. The storage medium can also include a combination of the above-mentioned storage media.

[0123] Although the embodiments of the present application are described in conjunction with the accompanying drawings, various modifications and changes can be made by those skilled in the art without departing from the spirit and scope of the present application, and such modifications and changes are intended to fall within the scope of the appended claims.

Claims

1. A method for human body recognition from a top-view perspective, characterized in that, include: Acquire the target image; The target image is identified based on the bounding box prediction head and the point prediction head in the preset recognition model, and the human detection bounding box and head vertex coordinates of the feature maps at multiple scales in the preset recognition model are obtained. The human detection bounding box based on multi-scale feature maps processes the small-scale feature maps in the multi-scale feature maps to obtain the target small-scale feature map. The target small-scale feature map is processed based on the appearance representation prediction head in the preset recognition model to obtain the target representation vector for human body recognition. The preset recognition model is trained according to the following method: A labeled image is acquired, comprising detection boxes, head vertices, and identifiers for each target object. The labeled image is then processed using box prediction heads and point prediction heads. The box prediction heads obtain box coordinates and box confidence scores, while the point prediction heads obtain head vertex coordinates and head vertex confidence scores. A small-scale feature image is processed based on the box coordinates to obtain a target feature image. The target feature image is cropped to obtain a target region, which is then input into an appearance representation prediction head to obtain an appearance representation vector. A target loss function is constructed based on the box coordinates, box confidence scores, head vertex coordinates, head vertex confidence scores, and appearance representation vector, and this target loss function is optimized to determine the preset recognition model. The bounding box confidence score includes bounding box category confidence score and bounding box target confidence score. The construction of the target loss function based on the bounding box coordinates, bounding box confidence score, head vertex coordinates, head vertex confidence score, and appearance representation vector includes: Calculate the losses based on the bounding box coordinates, bounding box confidence, top vertex coordinates, top vertex confidence, and appearance representation vector respectively to obtain the bounding box coordinate loss, bounding box target confidence loss, bounding box category confidence loss, top vertex coordinate loss, top vertex confidence loss, and appearance representation vector loss. The target loss function is determined according to the following formula: Among them, L total Let M represent the target loss function, and M represent the number of scales. (j = a, b, c, d, e, f) represent the bounding box coordinate loss, bounding box target confidence loss, bounding box category confidence loss, top vertex confidence loss, top vertex coordinate loss, and appearance representation vector loss, respectively. This represents the weighted parameters associated with each loss.

2. The method according to claim 1, characterized in that, The multi-scale feature maps include large-scale feature maps, medium-scale feature maps, and small-scale feature maps; the human detection bounding box based on the multi-scale feature maps and the head vertex coordinates are used to process the small-scale feature maps in the multi-scale feature maps to obtain the target small-scale feature map, including: The human detection boxes in the large-scale feature map and the medium-scale feature map are mapped to the small-scale feature map to obtain the target small-scale feature map.

3. The method according to claim 2, characterized in that, The large-scale feature map, medium-scale feature map, and small-scale feature map each include a bounding box prediction head and a point prediction head; the target image is identified based on the bounding box prediction head and point prediction head in the preset recognition model to obtain the human detection bounding box and head vertex coordinates of the feature maps at various scales in the preset recognition model, including: Large-scale, medium-scale, and small-scale feature images are processed based on the respective bounding box prediction heads and point prediction heads to obtain human detection bounding boxes, head vertex coordinates, bounding box confidence, and head vertex confidence for feature maps of various scales.

4. The method according to claim 1, characterized in that, The step of processing the small-scale feature map of the target based on the appearance representation prediction head in the preset recognition model to obtain the target representation vector includes: Cropping the small-scale feature map of the target yields the target region image; The target region image is identified based on the appearance representation prediction head in the preset recognition model to obtain the target representation vector.

5. The method according to claim 1, characterized in that, The preset recognition model includes large-scale feature images, medium-scale feature images, and small-scale feature images. The process of processing the small-scale feature images based on the coordinates of each bounding box to obtain the target feature image includes: The bounding box coordinates in the large-scale feature image and the bounding box coordinates in the medium-scale feature image are respectively mapped to the small-scale feature image to obtain the target feature image.

6. A human body recognition device for a top-view perspective, characterized in that, The device includes: The image acquisition module is used to acquire the target image; The image recognition module is used to recognize the target image based on the bounding box prediction head and the point prediction head in the preset recognition model, and to obtain the human detection bounding box and head vertex coordinates of the feature maps at multiple scales in the preset recognition model. The feature processing module is used to process the small-scale feature maps in the multi-scale feature maps based on the human detection box in the multi-scale feature maps to obtain the target small-scale feature map. The representation processing module is used to process the small-scale feature map of the target based on the appearance representation prediction head in the preset recognition model to obtain the target representation vector for human body recognition. The device further includes: The labeled image acquisition module is used to acquire labeled images, which include detection boxes, head vertices, and identifiers for each target object; The labeled image processing module is used to process the labeled image based on each bounding box prediction head and point prediction head. The bounding box prediction head is processed to obtain the bounding box coordinates and bounding box confidence, and the point prediction head is processed to obtain the head vertex coordinates and head vertex confidence. The target feature image acquisition module is used to process the small-scale feature image based on the coordinates of each bounding box to obtain the target feature image; The region cropping module is used to crop the target feature image to obtain the target region, and input the target region into the appearance representation prediction head to obtain the appearance representation vector. The model determination module is used to construct a target loss function based on the bounding box coordinates, bounding box confidence, head vertex coordinates, head vertex confidence, and appearance representation vector, and to optimize the target loss function to determine the preset recognition model; The bounding box confidence score includes the bounding box category confidence score and the bounding box target confidence score. The model determination module includes: The loss calculation unit is used to calculate the losses based on the bounding box coordinates, bounding box confidence, top vertex coordinates, top vertex confidence, and appearance representation vector, respectively, to obtain the bounding box coordinate loss, bounding box target confidence loss, bounding box category confidence loss, top vertex coordinate loss, top vertex confidence loss, and appearance representation vector loss. The target loss function is determined according to the following formula: Among them, L total Let M represent the target loss function, and M represent the number of scales. (j = a, b, c, d, e, f) represent the bounding box coordinate loss, bounding box target confidence loss, bounding box category confidence loss, top vertex confidence loss, top vertex coordinate loss, and appearance representation vector loss, respectively. This represents the weighted parameters associated with each loss.

7. An electronic device, characterized in that, include: A memory and a processor are communicatively connected, the memory storing computer instructions, and the processor executing the computer instructions to perform the human body recognition method for a top-view perspective as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the human body recognition method for a top-view perspective as described in any one of claims 1-5.

Images