CN116863509B - Method for human silhouette detection and gesture recognition using improved PolarMask - Google Patents

Abstract

The present invention uses an improved PolarMask method for humanoid contour detection and gesture recognition, adopts an improved PolarMask model, and designs humanoid contour polar coordinate modeling based on humanoid contour characteristics; then constructs an improved PolarMask model as a humanoid contour segmentation model, and adds The channel attention mechanism module adds skip connections to the original YOLOV7-based feature pyramid network to make up for the loss of detailed information during the feature fusion process. Finally, a training strategy based on weak labels is adopted to train a model that can identify features containing A primary humanoid outline segmentation model based on the rectangular frame of humanoid position information and the human posture type; in the formal training process, the pre-trained weights that have learned the relevant information of the humanoid outline in advance are used for transfer learning, and in the real humanoid In the process of contour learning, the predicted humanoid contours are continuously converged, and the humanoid contours and posture types are accurately identified.

Description

Method for human silhouette detection and gesture recognition using improved PolarMask

Technical field

The invention belongs to the field of image recognition, and in particular relates to a method of using an improved PolarMask for humanoid contour detection and gesture recognition.

Background technique

With the development of domestic computer technology and imaging technology, humanoid contour detection and gesture recognition methods have received more and more attention. Various fields have put forward higher demands for humanoid contour recognition and posture recognition. Humanoid contour detection and posture recognition technology can play a role in improving human-computer interaction, motion analysis, health monitoring, virtual reality and augmented reality, and security monitoring. play a key role. They provide more intuitive, efficient and immersive solutions for various application scenarios, and promote the development of artificial intelligence technology in practical applications. Especially for the needs of human body behavior recognition and intelligent monitoring, most of the human body behavior recognition and monitoring in existing video images are based on target recognition methods, which roughly identify the person's location and the state of the human body at that time, including upright or fallen. . Through the method of target recognition and posture recognition, first obtain the target frame containing the human figure, and identify the posture of the human body through the coordinates of the human skeleton nodes. Although the current method can achieve the effect of identifying the position and status of the human body to a certain extent, it cannot accurately identify the outline of the human body. Therefore, a humanoid outline recognition method is needed, which can not only accurately identify the human body, but also accurately segment the human body outline, provide more useful information for judging the human body's posture, thereby improving the accuracy of human body posture recognition, and at the same time, it is more conducive to Apply this technology to more related fields.

The accuracy of the instance segmentation method based on PolarMask is much lower than other instance segmentation methods such as Mask R-CNN. And because it is a deep learning model, PolarMask requires a large number of data sets for training to achieve the expected results. Labeling a picture by segmentation type is very time-consuming and laborious. It takes 1 minute on average to label a picture, so training Models also require a lot of investment.

Contents of the invention

The purpose of the present invention is to provide a method for humanoid contour detection and posture recognition using improved PolarMask to achieve the purpose of human body contour segmentation and posture recognition. This method can not only greatly reduce costs, but also save training time. , speed up reasoning.

The present invention uses an improved PolarMask method for humanoid contour detection and gesture recognition, adopts an improved PolarMask model, and uses the ratio of the length and width of the humanoid bounding box identified by the Polarmask model to allocate each area in the bounding box based on the characteristics of the humanoid outline. The number of rays is used to design the humanoid contour polar coordinate modeling; then an improved PolarMask model is constructed as a humanoid contour segmentation model. Before the feature pyramid network performs feature fusion, a channel attention mechanism module is added after each feature of different scales. Skip connections were added to the original feature pyramid network based on YOLOV7 to make up for the loss of detailed information during the feature fusion process. Finally, a training strategy based on weak labels was adopted, and the Box type weak label data set was used to pre-program the humanoid contour segmentation model. Training is used to train a primary humanoid contour segmentation model that can identify rectangular boxes containing humanoid position information and human posture types; in the formal training process, the pre-trained model that has learned the relevant information of the humanoid contour in advance is used. The training weights are used for transfer learning. In the process of learning the real humanoid outline, the predicted humanoid outline is continuously converged and the humanoid outline and posture type are accurately identified;

The constructed and improved PolarMask model is used as a humanoid contour segmentation model. Based on the original PolarMask model and the network structure of YOLOV7, the YOLOAT_FPN feature pyramid network is designed to replace the FPN network structure of the original PolarMask model and improve the backbone of the original PolarMask model. Network and feature pyramid structure; human silhouette segmentation model consists of one encoder and three decoders;

The encoder uses the YOLOAT_FPN feature pyramid network, which is improved as follows based on the backbone network of YOLOV7:

(1) Replaced the activation function of the original convolution module and replaced the original activation function SiLU with a nonlinear activation function GELU used in natural language processing;

(2) The channel attention mechanism module is added before feature fusion in the feature pyramid;

(3) Use the multi-scale feature maps extracted from the original YOLOV7 backbone network to further extract important detailed information through the channel attention mechanism module, and jump-connect shallow information and deep information through a 1×1 convolution kernel. , completes the detailed information lost through feature fusion;

The three decoders refer to three parallel processing branches, namely the classification branch, the centrality branch and the polar coordinate mask branch. Among them, the classification branch uses 4×4 Conv and 1×1 Conv in turn to perform feature extraction. Extract and generate H×W×N feature maps to predict N postures to achieve prediction of segmented target categories. H and W represent the length and width of the input feature map respectively, and N represents the type of posture that needs to be predicted; center The degree branch uses the 4×4 Conv and 1×1 Conv shared with the classification branch to extract features, and generates an H×W×1 feature map to predict the polar coordinate center point; the polar coordinate mask branch uses 4× Conv of 4 and Conv of 1×1 are used for feature extraction, and a feature map of H×W×60 is generated to predict the distance of 60 rays in polar coordinates.

The improved PolarMask model is used. Based on the characteristics of the humanoid outline, the ratio of the length and width of the humanoid bounding box identified by the Polarmask model is used to allocate the number of rays in each area of the bounding box, and the design of polar coordinate modeling of the humanoid outline is carried out. Specifically: the four vertices A, B, C, and D of the humanoid bounding box recognized by the Polarmask model and the human body center O form four regions, and the proportion of each area is allocated according to the ratio of the length and width of the identified bounding box. The number of rays is used to design the humanoid contour polar coordinate modeling. The calculation formula is as shown in formula (2):

Among them, 0 is the center point of the humanoid. The four vertices A, B, C, and D of the humanoid bounding box and the humanoid center point O constitute four areas, which are the AOB area, BOC area, COD area, and AOD area. The Number _AOB represents The number of rays required to construct the human body outline in the AOB area. The Number _COD represents the number of rays required to construct the human body outline in the COD area. The Number _AOD represents the number of rays required to construct the human body outline in the AOD area. The Number _BOC represents the number of rays required to construct the human body outline in the BOC area; N represents the total number of rays; Y represents the height of the bounding box; X represents the width of the bounding box.

The channel attention mechanism module selects the SENet model. The SENet model includes two stages: compression and excitation. In the compression stage, the global spatial information is compressed, and then feature learning is performed in the channel dimension to form the attention weight of each channel. Finally, in the excitation stage, the attention weight generated in the compression stage is applied to the corresponding channel, specifically:

The compression stage is performed first, using Global pooling to compress the H×W×C input into an output of 1×1×C, followed by the excitation stage, which includes two fully connected layers. The first fully connected layer has C/r neurons. element, the output is 1×1×(C/r), and uses the activation function ReLU; the second fully connected layer has C neurons, restores the output to 1×1×C, and uses the activation function Sigmoid, where r is the compression value of the first fully connected layer; in the excitation stage, by learning the characteristic information of each channel, the attention weight of each channel is generated, and the final output channel attention weight of 1×1×C is compared with the original The corresponding channels of the feature map are multiplied.

A device that uses an improved PolarMask for humanoid contour detection and gesture recognition. The device includes a processor and a memory; the memory is used to store a computer program; the processor is used to execute any of the above improvements according to the computer program. PolarMask is a method for human silhouette detection and gesture recognition.

A computer-readable storage medium. The computer-readable storage medium is used to store a computer program. The computer program is used to execute any of the above methods for humanoid contour detection and gesture recognition using improved PolarMask.

A chip that runs instructions, and the chip is used to execute any of the above-mentioned methods of humanoid contour detection and gesture recognition using improved PolarMask.

The improved PolarMask model of the present invention draws on the network structure of YOLOV7 based on the original PolarMask model, adds jump connections to the backbone network, adds an attention mechanism module, and applies the improved PolarMask to humanoid contour instance segmentation and gesture recognition. Box type weak labels are used for pre-training of the improved PolarMask model, thus introducing transfer learning into the contour segmentation model. Compared with the existing technology, there are the following technical effects:

(1) The present invention segments the human body outline based on the improved PolarMask model, and assists in identifying the posture of the human body by calculating the distance and angle of 60 rays, and uses a weak label data set with a simple box type to perform the improved PolarMask model. Pre-training, and then perform transfer learning on the obtained pre-trained weights. If COCO type instance segmentation labels are used for pre-training based on the PolarMask model, the processing of the data set is very time-consuming, which will result in higher costs. The present invention uses Box type weak label data sets for model pre-training. The difficulty and time spent on labeling through Box type weak labels are much less than the performance of the instance segmentation type, so the required costs will also be reduced. . After using the weak label data set for model pre-training, you only need to use a small number of instance segmentation labels and make corrections to achieve better results. Compared with other methods based on PolarMask, this invention can save costs to a great extent.

(2) The present invention redesigns the polar coordinate modeling method, and changes the originally equally spaced rays in PolarMask into unevenly arranged rays according to the characteristics of the human form. Use more rays to represent complex contours and fewer rays to represent simple contours. It solves the redundancy phenomenon in the original PolarMask model when using uniformly distributed rays to represent the contour. It can finally use fewer rays to represent the contour of the human body, and can reduce the parameters of the model to obtain a more accurate human contour.

(3) This invention replaces the backbone network in the original PolarMask model and helps the model extract features more effectively. The new backbone network draws on the network structure of YOLOV7, improves the activation function of the convolution module, and adds skip connections to improve the network's ability to extract detailed information.

(4) The present invention introduces an attention mechanism module into the human contour segmentation model and uses a channel attention network before feature fusion, so that the model pays more attention to important feature information and can improve the segmentation accuracy of the network to a certain extent.

(5) The present invention adopts transfer learning. The value of pre-training weights can be closer to the optimal convergence point. Therefore, in the subsequent formal training process, it only takes shorter training time to reach the convergence point, making the network easier. Converging to the optimal point not only improves accuracy, but also improves efficiency and improves the generalization ability of the model to a certain extent.

Description of the drawings

Figure 1 is a model structure diagram of the present invention;

Figure 2 is an improved diagram of the convolution module of the present invention;

Figure 3 is an improved diagram of polar coordinate modeling based on humanoid characteristics of the present invention;

Figure 4 is a flow chart of the weak label-based migration training method of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only some of the embodiments of the present invention, not all of them. . Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

Embodiment 1

Embodiment 1 of the present invention relates to a method of using improved PolarMask for humanoid contour detection and gesture recognition. It adopts the polar coordinate modeling method of the improved PolarMask model and combines the characteristics of the humanoid to redesign the outline of the human body; and then uses the YOLOV7 backbone network As the basis, the backbone network of the original PolarMask model was re-improved, and the original convolution module was improved. Before feature fusion in the feature pyramid network, an attention mechanism module was added after each feature of different scales. In the original feature pyramid based on YOLOV7 Skip connections are added to the network to make up for the loss of detailed information in the feature fusion process. Finally, a training strategy based on weak labels is adopted, and weak label data sets are used to pre-train the humanoid contour segmentation model. During the formal training process, using The pre-trained weights are transferred to the learning process. In the process of learning the real humanoid outline, the predicted humanoid outline is continuously converged and the humanoid outline and posture type are accurately identified. The specific steps include the following steps:

Step 1. Construct an improved PolarMask model as a humanoid contour segmentation model

As shown in Figure 1, based on the original PolarMask model and the network structure of YOLOV7, the backbone network and feature pyramid structure of the original PolarMask model were improved, and the YOLOAT_FPN feature pyramid network was designed to replace the FPN network structure of the original PolarMask model; humanoid The contour segmentation model consists of an encoder and three decoders;

The encoder uses the YOLOAT_FPN feature pyramid network, which is improved as follows based on the backbone network of YOLOV7:

(1) The activation function of the original convolution module is replaced, and the original activation function SiLU is replaced by a nonlinear activation function GELU (Gaussian Error Linear Unit) used in natural language processing, as shown in Figure 2; the activation function GELU is A smooth nonlinear function with continuously differentiable properties can better adapt to the gradient descent algorithm and is easier to converge during the training process. The specific mathematical formula of the GELU activation function is shown in formula (1):

GELU(x)＝0.5×x×(1+tanh(sqrt(2/pi)×(x+0.044715×x^3))) (1)

(2) A channel attention mechanism module is added before feature fusion in the feature pyramid to help the model pay more attention to important information and ignore unimportant information, thereby improving the segmentation accuracy of the image;

(3) Improve the feature pyramid structure, add skip connections, and design the YOLOAT_FPN feature pyramid network to help the model improve the extraction of detailed information. Although the original YOLOV7 network structure performs multi-scale fusion through feature pyramid, in the process of feature fusion, as the number of network layers deepens, detailed features will inevitably be lost, and shallow networks will also be ignored. Regarding the importance of model identification, the present invention further extracts important detailed information from the feature maps of multiple scales extracted from the original YOLOV7 backbone network through the channel attention mechanism module (SENet model), and passes the 1×1 convolution kernel Jump connections are made between shallow information and deep information, as shown by the red dotted line in Figure 1, which to a large extent complements the detailed information lost through feature fusion.

The three decoders refer to three parallel processing branches, namely the classification branch, the centrality branch and the polar coordinate mask branch. Among them, the classification branch uses 4×4 Conv and 1×1 Conv in turn to perform feature extraction. Extract and generate H×W×N feature maps to predict N postures to achieve prediction of segmented target categories. H and W represent the length and width of the input feature map respectively, and N represents the type of posture that needs to be predicted; center The degree branch uses the 4×4 Conv and 1×1 Conv shared with the classification branch to extract features, and generates an H×W×1 feature map to predict the polar coordinate center point; the polar coordinate mask branch uses 4× Conv of 4 and Conv of 1×1 are used for feature extraction, and a feature map of H×W×60 is generated to predict the distance of 60 rays in polar coordinates. At the same time, the present invention redesigns the arrangement of these 60 rays according to the characteristics of the humanoid. method, making it more suitable for segmentation of human images. The specific method will be introduced in detail in step 2.

The attention mechanism module selects the SENet model (Squeeze-and-Excitation Networks compression and excitation network) that belongs to channel attention. The SENet model includes two stages of compression and excitation. In the compression stage, the global spatial information is compressed, and then Feature learning is performed in the channel dimension to form the attention weight of each channel. Finally, in the excitation stage, the attention weight generated in the compression stage is applied to the corresponding channel, specifically:

The compression stage is performed first, using Global pooling to compress the H×W×C input into an output of 1×1×C, followed by the excitation stage, which includes two fully connected layers. The first fully connected layer has C/r neurons. element, the output is 1×1×(C/r), and uses the activation function ReLU; the second fully connected layer has C neurons, restores the output to 1×1×C, and uses the activation function Sigmoid, where r It is the compression value of the first fully connected layer. Generally, 16 has better effect. In the excitation stage, by learning the feature information of each channel, the attention weight of each channel is generated, and the final output channel attention weight of 1×1×C is multiplied by the channel corresponding to the original feature map.

Step 2. Based on the characteristics of the humanoid outline, the four vertices A, B, C, and D of the humanoid bounding box recognized by the Polarmask model and the human body center O form four regions, based on the ratio of the length and width of the identified bounding box. Allocate the number of rays to each area and perform polar coordinate modeling of the humanoid outline design

The original Polarmask model emits N different rays at equal intervals from the center of the object that needs to be segmented. By connecting the endpoints of the different rays in sequence, the outline of the object that needs to be segmented is designed, as shown in a) in Figure 3. This contour modeling method is more suitable for objects with circular structures. If this method is directly applied to the design of humanoid contours, there will be a lot of redundancy, which will not only increase unnecessary computational burden, but also fail to segment finer contours. This is because the contours of some parts of the humanoid only require a small number of rays. Line segments can well express the outline, but the outlines of some complex parts require more rays to express the outline.

As shown in b) in Figure 3, the number of rays in each area is allocated according to the ratio of the length and width of the humanoid bounding box recognized by the Polarmask model. The specific calculation formula is as shown in formula (2):

Among them, 0 is the center point of the humanoid. The four vertices A, B, C, and D of the humanoid bounding box and the humanoid center point O constitute four areas, which are the AOB area, BOC area, COD area, and AOD area. The Number _AOB represents As shown in b) of Figure 3, the number of rays required to construct the human body outline in the AOB area, the Number _COD· represents the number of rays required to construct the human body outline in the COD area; the Number _AOD represents the number of rays required to construct the human body outline in the AOD area The number of rays required; the Number _BOC represents the number of rays required to construct the human body outline in the BOC area; N represents the total number of rays; Y represents the height of the bounding box; X represents the width of the bounding box.

This improved humanoid contour polar coordinate modeling method can not only represent the contour in more detail and reduce redundancy, but also use this uneven polar coordinate modeling method to reduce the parameters of the model. Through testing, it was found that 90 uniform polar coordinates were originally needed. Only distributed rays can accurately represent the human silhouette. The improved humanoid contour modeling method of the present invention can well represent the humanoid contour using only 60 rays.

Step 3. Use the Box type weak label data set to pre-train the humanoid contour segmentation model to train a primary humanoid contour segmentation model that can identify rectangular boxes containing humanoid position information and human posture types; in formal training In the process, the pre-trained weights that have learned the relevant information of the humanoid outline in advance are used for transfer learning, and finally the humanoid outline and posture type are accurately identified;

Since the improved Polarmask model of the present invention needs to predict the humanoid bounding box and center point first during the training process, and allocate the number of rays to each area in the image according to the predicted bounding box and center point positions, so this model is not suitable for the bounding box and center point. The prediction is particularly important. Therefore, in order to predict the humanoid outline faster and more accurately, the present invention adopts a transfer learning method based on weak labels, and uses pre-training weights to learn the relevant information of the humanoid outline in advance, which is more conducive to the connection. The segmentation model converges to the optimal point, thereby improving the segmentation accuracy of the model. As shown in Figure 4.

Since some pictures contain multiple figures and overlap, if real tags are used directly, a lot of manpower will be consumed in the process of making and collecting tags. The present invention uses a Box type weak label data set to pre-train the humanoid contour segmentation model. Since most Box tags are weak tags in the VOC format, it is necessary to first convert the weak tags in the VOC format into tags in the COCO format to obtain a tag that can be provided to the PolarMask model for preliminary segmentation contours. The specific conversion method is: convert the VOC format to a tag The two points used to represent the rectangular frame of the recognition target, namely the point minimum in the upper left corner of the boxes and the point maximum in the lower right corner of the boxes, are converted into 4 "segmentation" in the COCO format.

The present invention uses the pre-training weights trained by the Box type weak label data set to perform migration learning, which can effectively improve the segmentation accuracy of the model. This data set has a large number. Compared with segmentation type masks, it is easier to use bounding boxes to annotate humanoid outlines. Before using real labels, the present invention has obtained a relatively accurate prediction of the rectangular frame. During the training process, the humanoid outline can converge correctly, and finally accurately identifies the humanoid outline and human posture type.

Embodiment 2

Embodiment 2 of the present invention provides a device for humanoid contour detection and gesture recognition using an improved PolarMask. The electronic device may be a terminal device or a server, or may be connected to other terminal devices or servers to implement the method of Embodiment 1 of the present invention. Terminal device or server.

The device may include: a processor (such as a CPU), a memory, and a data acquisition device; the processor connects to and controls the data acquisition device. Various instructions can be stored in the memory to complete various processing functions and implement the processing steps described in the method of the first embodiment.

Embodiment 3

Embodiment 3 of the present invention also provides a computer-readable storage medium. The computer-readable storage medium stores instructions, which when run on a computer, cause the computer to perform the processing steps described in the method of Embodiment 1 above.

Embodiment 4

Embodiment 4 of the present invention also provides a chip that runs instructions, and the chip is used to execute the processing steps described in the method of Embodiment 1.

Those skilled in the art should further realize that the units and algorithm steps of each example described in connection with the embodiments disclosed in the present invention can be implemented by electronic hardware, computer software, or a combination of both. In order to clearly illustrate the hardware and software Interchangeability, in the above description, the composition and steps of each example have been generally described according to functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Skilled artisans may implement the described functionality using different methods for each specific application, but such implementations should not be considered to be beyond the scope of the present invention.

The above-described specific embodiments further describe the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above-mentioned are only specific embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims (5)

1. The method of using improved PolarMask for humanoid contour detection and gesture recognition is characterized by: using the improved PolarMask model, based on the characteristics of humanoid contours, and assigning bounding boxes based on the ratio of the length and width of the humanoid bounding box identified by the PolarMask model. The number of rays in each area is used to design the humanoid contour polar coordinate modeling; then an improved PolarMask model is constructed as a humanoid contour segmentation model. Before the feature pyramid network performs feature fusion, channel attention is added after each feature of different scales. The force mechanism module adds skip connections to the original YOLOV7-based feature pyramid network to make up for the detailed information lost in the feature fusion process. Finally, a training strategy based on weak labels is adopted, and the Box type weak label data set is used to conduct humanoid contours The pre-training of the segmentation model is used to train a primary humanoid outline segmentation model that can recognize the rectangular frame containing humanoid position information and the human posture type; in the formal training process, the pre-trained humanoid outline is used to learn in advance The pre-trained weights of relevant information are transferred and learned. In the process of learning the real humanoid outline, the predicted humanoid outline is continuously converged and the humanoid outline and posture type are accurately identified; The constructed and improved PolarMask model is used as a humanoid contour segmentation model. Based on the original PolarMask model and the network structure of YOLOV7, the YOLOAT_FPN feature pyramid network is designed to replace the FPN network structure of the original PolarMask model and improve the backbone of the original PolarMask model. Network and feature pyramid structure; human silhouette segmentation model consists of one encoder and three decoders; The encoder uses the YOLOAT_FPN feature pyramid network, which is improved as follows based on the backbone network of YOLOV7: (1) Replaced the activation function of the original convolution module and replaced the original activation function SiLU with a nonlinear activation function GELU used in natural language processing; (2) A channel attention mechanism module is added before feature fusion in the feature pyramid; the channel attention mechanism module selects the SENet model. The SENet model includes two stages: compression and excitation. In the compression stage, the global spatial information is processed Compression, and then perform feature learning in the channel dimension to form the attention weight of each channel. Finally, in the excitation stage, the attention weight generated in the compression stage is applied to the corresponding channel, specifically: The compression stage is performed first, using Global pooling to compress the H×W×C input into an output of 1×1×C, followed by the excitation stage, which includes two fully connected layers. The first fully connected layer has C/r neurons. element, the output is 1×1×(C/r), and uses the activation function ReLU; the second fully connected layer has C neurons, restores the output to 1×1×C, and uses the activation function Sigmoid, where r is the compression value of the first fully connected layer; in the excitation stage, by learning the characteristic information of each channel, the attention weight of each channel is generated, and the final output channel attention weight of 1×1×C is compared with the original Multiply the corresponding channels of the feature map; (3) Use the multi-scale feature maps extracted from the original YOLOV7 backbone network to further extract important detailed information through the channel attention mechanism module, and jump-connect shallow information and deep information through a 1×1 convolution kernel. , completes the detailed information lost through feature fusion; The three decoders refer to three parallel processing branches, namely the classification branch, the centrality branch and the polar coordinate mask branch. Among them, the classification branch uses 4×4 Conv and 1×1 Conv in turn to perform feature extraction. Extract and generate H×W×N feature maps to predict N postures to achieve prediction of segmented target categories. H and W represent the length and width of the input feature map respectively, and N represents the type of posture that needs to be predicted; center The degree branch uses the 4×4 Conv and 1×1 Conv shared with the classification branch to extract features, and generates an H×W×1 feature map to predict the polar coordinate center point; the polar coordinate mask branch uses 4× Conv of 4 and Conv of 1×1 are used for feature extraction, and a feature map of H×W×60 is generated to predict the distance of 60 rays in polar coordinates. 2. The method for humanoid contour detection and gesture recognition using improved PolarMask according to claim 1, characterized in that the improved PolarMask model is used, and based on the humanoid contour characteristics, the length and width of the humanoid bounding box identified by the PolarMask model are To allocate the number of rays to each area in the bounding box with the ratio of the humanoid contour, the design of polar coordinate modeling of the humanoid outline is as follows: compare the four vertices A, B, C, and D of the humanoid bounding box recognized by the PolarMask model with the center of the human body O constitutes four areas, and the number of rays in each area is allocated according to the ratio of the length and width of the identified bounding box, and the design of polar coordinate modeling of human silhouette is carried out. The calculation formula is as shown in formula (2): Among them, O is the center point of the humanoid. The four vertices A, B, C, and D of the humanoid bounding box and the humanoid center point O form four areas, which are the AOB area, BOC area, COD area, and AOD area. The Number _AOB represents The number of rays required to construct the human body outline in the AOB area. The Number _COD represents the number of rays required to construct the human body outline in the COD area. The Number _AOD represents the number of rays required to construct the human body outline in the AOD area. The Number _BOC represents the number of rays required to construct the human body outline in the BOC area; N represents the total number of rays; Y represents the height of the bounding box; X represents the width of the bounding box. 3. A device using improved PolarMask for humanoid contour detection and gesture recognition, characterized in that: the device includes a processor and a memory; the memory is used to store a computer program; the processor is used to execute according to the computer program Any method of using the improved PolarMask for humanoid contour detection and gesture recognition described in claims 1-2. 4. A computer-readable storage medium, characterized in that: the computer-readable storage medium is used to store a computer program, and the computer program is used to execute any one of claims 1-2 using the improved PolarMask to perform humanoid processing. Methods for contour detection and gesture recognition. 5. A chip for running instructions, characterized in that: the chip is used to execute any of the methods of humanoid contour detection and gesture recognition using the improved PolarMask described in claims 1-2.