A gesture recognition model training method and gesture recognition method based on improved yolov5

Abstract

The invention discloses a gesture recognition model training method and a gesture recognition method based on improved yolov5, the training method comprising: constructing a sample data set; introducing an original yolov5 model, and improving the original yolov5 model to construct a gesture recognition model based on improved yolov5; inputting the sample data set into the gesture recognition model for training to obtain a gesture recognition model after training; wherein the original yolov5 model comprises a backbone network, a neck network and a detection network, and the backbone network, the neck network and the detection network are connected in sequence; improving the original yolov5 model, specifically comprising: adding a plurality of feature enhancement modules in the neck network, the feature enhancement module is used to provide different receptive fields for the input features by expanding the same convolution kernel with different expansion rates after receiving the input features. The invention helps to better fuse the features extracted by the neck network of the yolov5 model, thereby improving the target detection accuracy of the model and realizing more effective gesture recognition.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and more specifically, to a gesture recognition model training method and a gesture recognition method based on improved yolov5. Background Art

[0002] With the development of artificial intelligence, computer vision technology has gradually entered our daily lives, enriching people's spiritual and cultural life and bringing them a pleasant experience. The development of artificial intelligence interaction methods for gesture recognition allows users to use gestures to control or interact with devices, allowing computers to understand human behavior, thereby promoting the development of human-computer interaction.

[0003] At present, the gesture recognition technologies that have been implemented include those based on template matching and hidden Markov models, but these algorithms have complex modeling processes and poor generalization capabilities, making it difficult to effectively identify the type and state of gestures. At the same time, there is also a method based on hand key points for gesture recognition. This method requires the hand to be close to the camera so that the key points of the hand can be effectively located, and the state of the gesture can be judged based on the coordinate positions of the 21 key points of the hand. However, in the monitoring scenario, the hand is far away from the camera, the hand target is small, and the joints will be blurred. In this way, the use of hand key points for gesture recognition will result in inaccurate key point positioning, and gesture recognition cannot be effectively performed. Summary of the invention

[0004] The present invention aims to overcome at least one defect (shortcoming) of the above-mentioned prior art and provide a gesture recognition model training method and a gesture recognition method based on improved yolov5, so as to solve the problems of the existing gesture recognition technology having a complex modeling process, poor generalization ability, and inaccurate positioning of the hand key points due to the hand being far away from the camera and the hand target being small, thereby making it impossible to effectively perform gesture recognition.

[0005] The technical solution adopted by the present invention is a gesture recognition model training method based on improved yolov5, comprising the following steps:

[0006] Construct a sample dataset;

[0007] The original yolov5 model is introduced and improved to build a gesture recognition model based on the improved yolov5;

[0008] Inputting the sample data set into the gesture recognition model for training to obtain a trained gesture recognition model;

[0009] The original yolov5 model includes a backbone network, a neck network and a detection network, and the backbone network, the neck network and the detection network are connected in sequence;

[0010] The original yolov5 model is improved, specifically including: adding several feature enhancement modules in the neck network, wherein the feature enhancement modules are used to provide different receptive fields for the input features by expanding the same convolution kernel with different expansion rates after receiving the input features.

[0011] Furthermore, the feature enhancement module includes several parallel branches and an average pooling layer, each of the parallel branches is used to expand the input features of the same convolution kernel with different expansion rates, and the average pooling layer is used to fuse the features obtained after the expansion of each parallel branch.

[0012] As an improvement, the present invention adds a feature extraction module to the neck network of the original yolov5 model, and provides receptive fields of different sizes for the input features by performing different degrees of dilated convolution on the input features, which helps to better fuse the features extracted by the neck network of the yolov5 model, thereby improving the target detection accuracy of the model and achieving effective gesture recognition. Specifically, the feature extraction module added in the present invention is composed of a plurality of parallel branches with different dilation rates and an average pooling layer. The input features of the same convolution kernel first implement different degrees of dilated convolution through each parallel branch, providing receptive fields of different sizes for the input features, and then continue to fuse the features extracted after the dilated convolution of each parallel branch through the average pooling layer, thereby improving the model target detection accuracy and achieving effective gesture recognition.

[0013] Further, the original yolov5 model is improved, and further includes: adding a lightweight feature extraction network to the original yolov5 model, and adding several feature fusion modules;

[0014] The backbone network includes several outputs, the lightweight feature extraction network includes several outputs, each output of the lightweight feature extraction network and each output of the backbone network are respectively input to the feature fusion module for feature fusion, and the feature fusion module outputs the fused features to the neck network;

[0015] The output of the lightweight feature extraction network, the output of the backbone network and the feature fusion module are in a one-to-one relationship.

[0016] In order to better achieve the gesture recognition effect, the present invention also adds a lightweight feature extraction network and a feature fusion module to the original yolov5 model. The input sample data set is first subjected to feature extraction through the backbone network and the lightweight feature extraction network respectively, and then the features extracted by the lightweight feature extraction network and the features extracted by the backbone network are correspondingly fused through the feature fusion module, so that the improved model can extract features at different levels from multiple dimensions, and correspondingly fuse features at multiple levels extracted from multiple dimensions, thereby enriching the features extracted by the model and helping the model to extract more effective and comprehensive features.

[0017] Furthermore, the original yolov5 model is improved, and also includes: adding a plurality of detection heads in the detection network so that the number of the detection heads matches the number of the feature fusion modules.

[0018] In the original yolov5 model, since the number of features extracted by its backbone network and neck network is small, the number of detection heads used to perform the final detection target is relatively small. However, in the monitoring scenario, the hand is far away from the camera and the hand target is small. The original yolov5 model cannot fully extract features of the detection target, so it cannot recognize small gesture targets well. In the present invention, the improvement of the yolov5 model also includes adding several detection heads to the detection network, so that the number of detection heads matches the number of feature fusion modules. At the same time, because the feature fusion modules in the present invention are one-to-one with the features extracted by the lightweight feature extraction network and the backbone network, the number of final detection heads is also matched with the number of features extracted by the lightweight feature extraction network and the backbone network, that is, the added detection heads in the present invention can fully match the features of different levels extracted by the lightweight feature extraction network and the backbone network, and realize the detection of more comprehensive features of different levels, thereby improving the recognition and detection effect of the model for smaller gesture targets.

[0019] Furthermore, the lightweight feature extraction network includes several convolution modules and pooling modules, and CRELU is used as an activation function between the convolution modules and the pooling modules.

[0020] In the present invention, CRELU is used as the activation function between the convolution module and the pooling module of the lightweight feature extraction network, which can effectively reduce the number of output channels of the convolution module, so that the number of required convolution kernels is also relatively reduced, thereby reducing the amount of network calculation while ensuring accuracy, significantly improving the network calculation rate, and making the lightweight feature extraction network truly meet the lightweight requirements.

[0021] Furthermore, a sample data set is constructed, including:

[0022] Collecting gesture recognition sample data, and labeling the gesture recognition sample data with gesture types to obtain an original sample data set;

[0023] Preprocessing the original sample data set to construct a sample data set;

[0024] The preprocessing includes translation, left-right mirroring and color gamut change;

[0025] The gesture types include a thumbs-up gesture and a non-thumbs-up gesture.

[0026] The gesture recognition model of the present invention mainly recognizes two gestures, namely, like and non-like. Firstly, because the specific application scenario of the present invention is mainly to express the degree of preference for a program by gesture, only two gestures, namely, like and non-like, are needed to complete human-computer interaction and achieve the effect of expressing the degree of preference for a program by gesture; secondly, if multiple gestures are used for expression, such as using finger changes to represent different gestures, the camera may only capture obvious gesture changes when the hand is close to the camera and the hand is clear. However, at a long distance such as monitoring, since the hand image itself accounts for a small proportion, when the finger changes represent different gestures, the hand image captured by the camera does not change significantly enough, so different gestures cannot be recognized. However, only two gestures, namely, like and non-like, are used. Since the hand posture changes are obviously different, the camera can still recognize and capture obvious gesture changes, thereby accurately performing gesture recognition. Among them, the specific definition of the like gesture is a gesture of four fingers clenching the palm, the thumb extending, and the thumb extending angle is 0-180°. In addition, in the present invention, when constructing a sample data set, preprocessing operations such as translation, left-right mirroring, and color gamut change are performed on the original sample data set. This operation can effectively expand the sample data, thereby enriching the training sample data set.

[0027] On the other hand, another technical solution adopted by the present invention is a gesture recognition method based on improved yolov5, comprising the following steps:

[0028] Collect gesture data to be recognized;

[0029] The gesture data to be recognized is input into the gesture recognition model trained by the above training method, and the gesture recognition result is output.

[0030] Furthermore, the gesture data to be recognized is input into the gesture recognition model trained by the above training method, and the gesture recognition result is output, which specifically includes:

[0031] Inputting the gesture data to be recognized into the gesture recognition model trained by the above training method to obtain a preliminary gesture recognition result;

[0032] The preliminary gesture recognition result is post-processed by non-maximum suppression to obtain a final gesture recognition result, and the final gesture recognition result is output.

[0033] In the specific recognition process, the gesture recognition model will first generate multiple detection frames and the confidence corresponding to each detection frame based on the gesture data of multiple detection heads, so as to perform specific gesture recognition. In order to improve the model recognition effect, remove recognition interference, and obtain more accurate recognition results, the present invention will also perform a post-processing operation of non-maximum suppression after obtaining the preliminary gesture recognition results, which is mainly used to remove duplicate detection frames. Only one detection frame with the highest confidence is retained for one gesture data to be recognized, thereby obtaining the most accurate gesture recognition results.

[0034] On the other hand, another technical solution adopted by the present invention is a gesture recognition model training system based on improved yolov5, comprising:

[0035] A sample data set construction module, used to construct a sample data set;

[0036] A model building module, used for introducing the original yolov5 model, improving the original yolov5 model, and building a gesture recognition model based on the improved yolov5;

[0037] The model training module is used to input the sample data set into the gesture recognition model for training to obtain a gesture recognition model after training.

[0038] On the other hand, another technical solution adopted by the present invention is a gesture recognition system based on improved yolov5, comprising:

[0039] A gesture data collection module to be recognized, used to collect gesture data to be recognized;

[0040] The recognition module is used to input the gesture data to be recognized into the gesture recognition model trained by the above training method, and output the gesture recognition result.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] 1. In the present invention, a feature extraction module is added to the neck network of the original yolov5 model. By performing different degrees of dilated convolution on the input features, different sizes of receptive fields are provided for the input features, which helps to better integrate the features extracted by the neck network of the yolov5 model, thereby improving the target detection accuracy of the model and realizing effective gesture recognition;

[0043] 2. In the present invention, a lightweight feature extraction network and a feature fusion module are added to the original yolov5 model, so that the improved model can extract features at different levels from multiple dimensions, and correspondingly fuse features at multiple levels extracted from multiple dimensions, thereby enriching the features extracted by the model and helping the model to extract more effective and comprehensive features;

[0044] 3. In the present invention, several detection heads are added to the detection network of the original yolov5 model, so that the number of detection heads matches the number of feature fusion modules, and a more comprehensive detection of features at different levels is achieved, thereby improving the recognition and detection effect of the model for smaller gesture targets;

[0045] 4. In the present invention, CRELU is used as the activation function between the convolution module and the pooling module of the lightweight feature extraction network, which reduces the amount of network calculation while ensuring accuracy, significantly improves the network calculation rate, and enables the lightweight feature extraction network to truly meet the lightweight requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a flow chart of the training method in Example 1 of the present invention.

[0047] Figure 2 This is a specific flow chart of step S1 of the training method in Example 1 of the present invention.

[0048] Figure 3 This is a structural diagram of a gesture recognition model in Example 1 of the present invention.

[0049] Figure 4 This is a structural diagram of the C3 module of the gesture recognition model in Example 1 of the present invention.

[0050] Figure 5 This is a structural diagram of the bottleneck module of the gesture recognition model in Example 1 of the present invention.

[0051] Figure 6 This is a structural diagram of the SPPF module of the gesture recognition model in Example 1 of the present invention.

[0052] Figure 7 This is a structural diagram of a feature enhancement module of a gesture recognition model in Example 1 of the present invention.

[0053] Figure 8 This is a structural diagram of a feature fusion module of a gesture recognition model in Example 1 of the present invention.

[0054] Fig. 9 4 is a CRELU structure diagram of the gesture recognition model in Example 1 of the present invention.

[0055] Fig.10This is a flow chart of the identification method in Example 2 of the present invention.

[0056] Fig.11 This is a specific flow chart of step S200 of the identification method in Embodiment 2 of the present invention.

[0057] Fig.12 This is a structural diagram of the gesture recognition model training system in Example 3 of the present invention.

[0058] Fig.13 This is a structural diagram of a gesture recognition system in Example 4 of the present invention.

[0059] Explanation of the reference numerals: sample data set construction module 100 , model construction module 200 , model training module 300 , gesture data collection module to be recognized 400 , recognition module 500 . DETAILED DESCRIPTION

[0060] The drawings of the present invention are only for illustrative purposes and should not be construed as limiting the present invention. In order to better illustrate the following embodiments, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product; it is understandable to those skilled in the art that some well-known structures and their descriptions in the drawings may be omitted.

[0061] Example 1

[0062] like Figure 1 As shown, this embodiment provides a gesture recognition model training method based on improved yolov5, comprising the following steps:

[0063] S1, build a sample data set;

[0064] S2, introduce the original yolov5 model, improve the original yolov5 model, and build a gesture recognition model based on the improved yolov5;

[0065] S3. Input the sample data set into the gesture recognition model for training to obtain a trained gesture recognition model.

[0066] The above steps are described in detail below.

[0067] like Figure 2 As shown, step S1 specifically includes:

[0068] S10, collecting gesture recognition sample data, and annotating the gesture recognition sample data with gesture types to obtain an original sample data set;

[0069] S20, preprocessing the original sample data set to construct a sample data set;

[0070] Preprocessing includes translation, left-right mirroring, and changing color gamut;

[0071] The gesture types include like gestures and not like gestures.

[0072] Specifically, due to the scarcity of sample data for thumbs-up gesture recognition, there is no public thumbs-up gesture recognition sample data set that meets the project requirements on the Internet. There are only some hand data marked with polygons. In order to make full use of this part of the data, in this embodiment, the marked polygonal box is converted into a rectangular box to facilitate subsequent use in target detection. At the same time, in order to collect thumbs-up gesture recognition sample data, this embodiment organizes personnel to simulate the thumbs-up action behavior when watching the program on site in multiple conference rooms to collect video data from the camera. In order to enrich the data samples, a mobile phone is used to move and shoot at multiple angles. The collected video data is extracted and saved every 5 frames. Before processing, the image can also be subtracted from the mean, divided by the variance, and other operations. Finally, the labelimg software is used to mark the gesture type. The gesture type is divided into two types: thumbs-up and non-thumbs-up. The thumbs-up gesture is specifically defined as a gesture of four fingers clenching the palm, the thumb extending, and the thumb extending angle is 0-180°.

[0073] Specifically, in step S2, the gesture recognition model structure based on the improved yolov5 is as follows: Figure 3 shown.

[0074] Among them, the original yolov5 model includes the backbone network ( Figure 3 Backbone module in the Figure 3 Neck module in) and the detection network ( Figure 3 The Head module in the network), the backbone network, the neck network and the detection network are connected in sequence; specifically, the backbone network includes multiple convolution modules, C3 modules and SPPF modules, and the neck structure also includes multiple C3 modules, wherein the specific structure of the C3 module is as follows Figure 4 As shown, C3 includes multiple bottleneck modules. The specific structure of the bottleneck module is as follows Figure 5 As shown, the specific structure of the SPPF module is as follows Figure 6 shown.

[0075] In step S2, the original yolov5 model is improved, specifically including: adding a feature enhancement module in the neck network ( Figure 3 The FA module in the neck network) is used to provide different receptive fields for the input features by expanding the same convolution kernel with different expansion rates after receiving the input features. Specifically, in this embodiment, a corresponding number of feature enhancement modules are added between the Upsample upsampling module in the neck network and the convolution module connected thereto.

[0076] Furthermore, in this embodiment, the specific structure of the feature enhancement module is as follows: Figure 7 As shown in Figure 1, it includes a convolution module with a convolution kernel of 3×3, four parallel branches with expansion rates of 1, 3, 5, and 7, and an average pooling layer ( Figure 4 The mean module in the convolution operation model is used to expand the input features of the same convolution kernel with different expansion rates, and the average pooling layer is used to fuse the features obtained after the expansion of each parallel branch. Each parallel branch includes a dilation module and a convolution module with a convolution kernel of 1×1. The dilation rates of each dilation module are 1, 3, 5, and 7 respectively. Specifically, the input features are first convolved through a convolution module with a convolution kernel of 3×3, and then expanded through the dilation modules of four parallel branches with dilation rates of 1, 3, 5, and 7 respectively, and then convolved through a convolution module with a convolution kernel of 1×1 respectively. Finally, an average pooling layer is used to fuse the features obtained after the expansion of each parallel branch, thereby providing receptive fields of different sizes for the input features, improving the model's target detection accuracy, and achieving effective gesture recognition.

[0077] Furthermore, in step S2, the original yolov5 model is improved, and further includes: adding a lightweight feature extraction network ( Figure 3 Minornetwork), and adding a feature fusion module between the backbone network and the feature extraction network ( Figure 3 FF module);

[0078] Specifically, in this embodiment, the backbone network includes four outputs, the lightweight feature extraction network includes four outputs, each output of the lightweight feature extraction network and each output of the backbone network are respectively input to the feature fusion module for feature fusion, and the feature fusion module outputs the fused features to the neck network;

[0079] The output of the lightweight feature extraction network, the output of the backbone network, and the feature fusion module are in a one-to-one relationship.

[0080] In this embodiment, the specific structure of the feature fusion module is as follows: Figure 8 As shown, specifically, the output of the backbone network and the output of the lightweight feature extraction network ( Figure 5 In the above example, input1 and input2 are firstly computed through a tensor product module, and then convolution operations are performed through two convolution modules with convolution kernels of 3×3 and 1×1 respectively, to obtain the feature fusion ( Figure 5 output).

[0081] Furthermore, in step S2, the original yolov5 model is improved, and further includes: adding detection heads to the detection network so that the number of detection heads matches the number of feature fusion modules.

[0082] Specifically, in this embodiment, the number of outputs of the backbone network and the lightweight feature extraction network are both four, and the output of the lightweight feature extraction network, the output of the backbone network and the feature fusion module are in a one-to-one relationship. Therefore, the number of feature fusion modules is also four, and the number of detection heads is also four. Compared with the original yolov5 model, a detection head is added to the shallow feature layer to detect smaller gesture targets.

[0083] Specifically, in this embodiment, the lightweight feature extraction network includes a combination of two groups of convolution modules and pooling modules connected in sequence and three separate convolution modules with a convolution kernel of 3×3, and also includes four convolution modules with a convolution kernel of 1×1, wherein the input feature passes through the combination of the two groups of convolution modules and the pooling module, and then passes through a convolution module with a convolution kernel of 1×1 to obtain the first output of the lightweight feature extraction network, and the input feature passes through the combination of the two groups of convolution modules and the pooling module, and then passes through the first convolution module with a convolution kernel of 3×3, and then passes through a convolution module with a convolution kernel of 1×1. The second output of the lightweight feature extraction network is obtained by a convolution module with a convolution kernel of 1×1. After the input features pass through the combination of two sets of convolution modules and pooling modules, they pass through two convolution modules with convolution kernels of 3×3 in sequence, and then pass through a convolution module with a convolution kernel of 1×1 to obtain the third output of the lightweight feature extraction network. After the input features pass through the combination of two sets of convolution modules and pooling modules, they pass through three convolution modules with convolution kernels of 3×3 in sequence, and then pass through a convolution module with a convolution kernel of 1×1 to obtain the fourth output of the lightweight feature extraction network. Among them, in the combination of two sets of convolution modules and pooling modules, CRELU is used as the activation function between the convolution module and the pooling module. Specifically, the specific structure of CRELU is as follows: Fig. 9 shown.

[0084] Example 2

[0085] like Fig.10 As shown, this embodiment provides a gesture recognition method based on improved yolov5, comprising the following steps:

[0086] S100, collecting gesture data to be recognized;

[0087] S200: Input the gesture data to be recognized into the gesture recognition model trained by the method described in Example 1, and output the gesture recognition result.

[0088] Further, such as Fig.11 As shown, step S200 specifically includes:

[0089] S210, inputting the gesture data to be recognized into the gesture recognition model trained by the method described in Example 1 to obtain a preliminary gesture recognition result;

[0090] S220 , performing non-maximum suppression post-processing on the preliminary gesture recognition result to obtain a final gesture recognition result, and outputting the final gesture recognition result.

[0091] Specifically, in this embodiment, after obtaining the final gesture recognition result, the number of specific like gestures can be further counted, thereby realizing not only the recognition application of the like gesture but also the counting application of the like gesture number.

[0092] Example 3

[0093] like Fig.12 As shown, this embodiment provides a gesture recognition model training system based on improved yolov5, including:

[0094] The sample data set construction module 100 is used to construct a sample data set;

[0095] The model building module 200 is used to introduce the original yolov5 model, improve the original yolov5 model, and build a gesture recognition model based on the improved yolov5;

[0096] The model training module 300 is used to input the sample data set into the gesture recognition model for training to obtain a trained gesture recognition model.

[0097] Example 4

[0098] like Fig.13 As shown, this embodiment provides a gesture recognition system based on improved yolov5, including:

[0099] The gesture data collection module 400 is used to collect gesture data to be recognized;

[0100] The recognition module 500 is used to input the gesture data to be recognized into the gesture recognition model trained by the method described in Example 1, and output the gesture recognition result.

[0101] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solution of the present invention, and are not intended to limit the specific implementation methods of the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the claims of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A gesture recognition model training method based on improved yolov5, characterized in that: The following steps are involved: Construct a sample dataset; The original yolov5 model is introduced and improved to build a gesture recognition model based on the improved yolov5; The original yolov5 model includes a backbone network, a neck network and a detection network, and the backbone network, the neck network and the detection network are connected in sequence; Add a lightweight feature extraction network to the original yolov5 model, and add several feature fusion modules; The backbone network includes several outputs, the lightweight feature extraction network includes several outputs, each output of the lightweight feature extraction network and each output of the backbone network are respectively input to the feature fusion module for feature fusion, and the feature fusion module outputs the fused features to the neck network; The output of the lightweight feature extraction network, the output of the backbone network and the feature fusion module are in a one-to-one relationship; Inputting the sample data set into the gesture recognition model for training to obtain a trained gesture recognition model; Adding a number of detection heads to the detection network so that the number of the detection heads matches the number of the feature fusion modules; The original yolov5 model is improved, specifically including: adding several feature enhancement modules in the neck network, wherein the feature enhancement modules are used to provide different receptive fields for the input features by expanding the same convolution kernel with different expansion rates after receiving the input features.

2. A gesture recognition model training method based on improved yolov5 according to claim 1, characterized in that: The feature enhancement module includes several parallel branches and an average pooling layer, each of the parallel branches is used to expand the input features of the same convolution kernel with different expansion rates, and the average pooling layer is used to fuse the features obtained after the expansion of each parallel branch.

3. The method for training a gesture recognition model based on improved yolov5 according to claim 1, characterized in that: The lightweight feature extraction network includes several convolution modules and pooling modules, and CRELU is used as an activation function between the convolution modules and the pooling modules.

4. A gesture recognition model training method based on improved yolov5 according to any one of claims 1 to 3, characterized in that: Construct a sample data set, including: Collecting gesture recognition sample data, and labeling the gesture recognition sample data with gesture types to obtain an original sample data set; Preprocessing the original sample data set to construct a sample data set; The preprocessing includes translation, left-right mirroring and color gamut change; The gesture types include a thumbs-up gesture and a non-thumbs-up gesture.

5. A gesture recognition method based on improved yolov5, characterized in that: The following steps are involved: Collect gesture data to be recognized; The gesture data to be recognized is input into the gesture recognition model trained by the method according to any one of claims 1 to 4, and the gesture recognition result is output.

6. A gesture recognition method based on improved yolov5 according to claim 5, characterized in that: Inputting the gesture data to be recognized into the gesture recognition model trained by the method according to any one of claims 1 to 4, and outputting the gesture recognition result, specifically comprising: Inputting the gesture data to be recognized into the gesture recognition model trained by the method according to any one of claims 1 to 4 to obtain a preliminary gesture recognition result; The preliminary gesture recognition result is post-processed by non-maximum suppression to obtain a final gesture recognition result, and the final gesture recognition result is output.

7. A gesture recognition model training system based on improved yolov5, characterized in that: The system implements the gesture recognition model training method based on improved yolov5 according to any one of claims 1 to 4, and the system comprises: A sample data set construction module, used to construct a sample data set; A model building module, used for introducing the original yolov5 model, improving the original yolov5 model, and building a gesture recognition model based on the improved yolov5; The model training module is used to input the sample data set into the gesture recognition model for training to obtain a gesture recognition model after training.

8. A gesture recognition system based on improved yolov5, characterized in that: include: A gesture data collection module to be recognized, used to collect gesture data to be recognized; The recognition module is used to input the gesture data to be recognized into the gesture recognition model trained by the method according to any one of claims 1 to 4, and output the gesture recognition result.

Images