A method for evaluating individual soldier tactical actions based on human posture estimation
By constructing a lightweight network model to evaluate individual soldier tactical actions, the problem of lack of objective evaluation criteria in existing technologies is solved, and efficient and accurate tactical action scoring and bone node estimation are achieved.
Patent Information
- Application Number
- CN202111658004.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-30
- Publication Date
- 2025-09-19
- Estimated Expiration
- 2041-12-30
AI Technical Summary
The existing individual soldier training system lacks objective and quantitative tactical action evaluation standards, which leads to subjectivity and inefficiency in human evaluation.
By collecting tactical movement information of soldiers during training in real time, using artificial intelligence and multi-dimensional human posture estimation methods, a lightweight network model is constructed for data analysis to achieve objective scoring of tactical movements.
It achieves objective quantitative evaluation of individual soldier tactical actions, improves evaluation efficiency and accuracy, reduces computational overhead, and supports multi-task rapid estimation of key human bone nodes.
Smart Images
Figure CN114399836B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to an action evaluation method, in particular to an individual soldier tactical action evaluation method based on human body posture estimation. Background Art
[0002] The existing individual soldier training system is mainly used for information management, but there is no set of objective and quantitative standards for evaluating training itself, especially training actions. Training officers often rely on their experience to make manual judgments, which are prone to subjective conclusions, lack of quantification, and other uncertainties. At the same time, traditional manual evaluation cannot replicate experience on a large scale, which also leads to low evaluation efficiency. Summary of the Invention
[0003] This invention aims to address the technical issues of existing individual soldier training processes, which lack universal standards for objectively quantifying individual soldier tactical movements, and the resulting subjective conclusions, lack of quantification, and low evaluation efficiency. Instead, it provides a method for evaluating individual soldier tactical movements based on human posture estimation. This method collects multi-dimensional information about soldiers' tactical movements during training and uses artificial intelligence to analyze the data to generate training scores.
[0004] The technical solution of the present invention is:
[0005] A method for evaluating individual soldier tactical actions based on human posture estimation comprises the following steps:
[0006] 1) Collect students’ body posture images in real time and identify target images;
[0007] 1.1) Establish the first backbone network model;
[0008] 1.2) Extract 8, 16, and 32 times downsampled features from the first backbone network model constructed in step 1.1);
[0009] 1.3) The features extracted from step 1.2) with 8, 16, and 32 times downsampling are input into the feature pyramid network for feature fusion;
[0010] 1.4) Outputting the fusion result of step 1.3) to the detection head to detect and identify the target image;
[0011] 2) Evaluate human posture;
[0012] 2.1) Construct a lightweight network; the lightweight network consists of six convolutional layers:
[0013] The first layer is the Conv / Relu layer, with a kernel size of 3×3, a stride of 2, and 32 channels;
[0014] The second layer is the Conv / Relu layer, with a core size of 3×3, a stride of 1, and 64 channels;
[0015] The third layer is the Conv / Relu layer, with a core size of 3×3, a stride of 2, and 64 channels;
[0016] The fourth layer is the Conv / Relu layer, with a core size of 3×3, a stride of 1, and 64 channels;
[0017] The fifth layer is the TransConv / BN / Relu layer with a kernel size of 2×2, a stride of 2, and 64 channels.
[0018] The sixth layer is the TransConv / BN / Relu layer with a kernel size of 2×2, a stride of 2, and a channel of 32.
[0019] 2.2) Transferring the latent pose from a pre-trained larger teacher model to a constructed lightweight network to form a student model through a fast pose distillation model training method based on heat maps of key human bone nodes;
[0020] 2.3) Downsample the target image detected and identified in step 1.4 twice through strided convolution;
[0021] 2.4) The latent posture learned in the teacher model is used to train the student model, and the target image obtained by twice downsampling in step 2.3) is transmitted to the student model to obtain the confidence of the target image; the confidence of the target image is then compared with a preset confidence standard value. If it is greater than or equal to the confidence standard value, an image display of the heat map of the key bone nodes of the human body is output, and then the scoring system is entered and the score is output; if it is less than the confidence standard value, the corresponding point deduction operation is performed, and then the scoring system is entered and the score is output.
[0022] Furthermore, in step 1.1), the first backbone network model includes:
[0023] The operator of the first stage is 3×3 convolution, the resolution is 224×224, the channel is 32, and the number of layers is 1;
[0024] The operator in the second stage is MBConv1, with k3×3, resolution of 112×112, channels of 16, and layers of 1;
[0025] The operator in the third stage is MBConv6, with k3×3, resolution of 112×112, channels of 24, and layers of 2;
[0026] The operator in the fourth stage is MBConv6, with k5×5, resolution 56×56, 40 channels, and 2 layers;
[0027] The operator of the fifth stage is MBConv6, with k3×3, resolution of 28×28, 80 channels, and 3 layers.
[0028] Furthermore, in step 1.1), the first backbone network model is composed of MBConv1 and MBConv6; the MBConv1 includes a 1×1 convolution, a batch normalization layer and an activation layer arranged in parallel, DWConvk×k, a batch normalization layer arranged in parallel, an activation layer and a compression and excitation module, a 1×1 convolution, a batch normalization layer and Add, and a side connection between the input end and Add; wherein the step size of DWConvk×k is 1; the MBConv6 includes a 1×1 convolution, a batch normalization layer and an activation layer arranged in parallel, DWConvk×k, a batch normalization layer arranged in parallel, an activation layer and a compression and excitation module, a 1×1 convolution and a batch normalization layer, connected in sequence from the input end to the output end; wherein the step size of DWConv k×k is 2.
[0029] Furthermore, in step 1.4), the detection head adopts a Focal loss function.
[0030] Furthermore, in step 1.4), the Focal loss function adopted by the detection head is a distributed Focal loss function.
[0031] Furthermore, in step 2.4), the confidence standard value is 0.85±0.02.
[0032] Beneficial effects of the present invention:
[0033] 1. The present invention collects tactical action information of soldiers during training in multiple dimensions and uses artificial intelligence to perform data analysis to obtain the soldier training score.
[0034] 2. The present invention designs an efficient single-soldier human body detection backbone network model, which can balance the scaling of three dimensions and improve the efficiency of the network model without adding additional computing overhead.
[0035] 3. The present invention is intelligent and efficient, and can realize multi-task rapid estimation of key bone nodes of the human body. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 It is the MBConv1 flow chart of the present invention;
[0037] Figure 2 It is the MBConv6 flow chart of the present invention;
[0038] Figure 3 It is a flow chart of human posture assessment in the present invention. DETAILED DESCRIPTION
[0039] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments and drawings.
[0040] The present invention provides a method for judging individual soldier tactical actions based on human body posture estimation, comprising the following steps:
[0041] 1) Collect students’ body posture images in real time and identify target images;
[0042] 1.1) Establish the first backbone network model;
[0043] The first backbone network model is composed of MBConv1 and MBConv6. Figure 1 As shown in , MBConv1 includes 1×1 convolution, batch normalization layer and activation layer set in parallel, DWConvk×k, batch normalization layer set in parallel, activation layer and compression and excitation module, 1×1 convolution, batch normalization layer and Add (fusion module) connected in sequence from input to output, and side connection from input to Add (fusion module); wherein, the stride of DWConvk×k is 1; as shown in Figure 2 As shown in the figure, MBConv6 includes a 1×1 convolution, a batch normalization layer and an activation layer set in parallel, DWConvk×k, a batch normalization layer set in parallel, an activation layer and a compression and excitation module, a 1×1 convolution and a batch normalization layer connected in sequence from the input to the output; among them, the stride of DWConvk×k is 2.
[0044] We can also optimize the first backbone network model to obtain a simpler one. That is, remove all the original convolutional layers in the first backbone network model and only retain the 1×1 convolution extracted from the first backbone network model to align the feature dimensions and channels. Upsampling and downsampling are both completed by interpolation. Finally, the simplified first backbone network model is as follows:
[0045] The operator of the first stage is 3×3 convolution, the resolution is 224×224, the channel is 32, and the number of layers is 1;
[0046] The operator in the second stage is MBConv1, with k3×3, resolution of 112×112, channels of 16, and layers of 1;
[0047] The operator in the third stage is MBConv6, with k3×3, resolution of 112×112, channels of 24, and layers of 2;
[0048] The operator in the fourth stage is MBConv6, with k5×5, resolution 56×56, 40 channels, and 2 layers;
[0049] The operator in the fifth stage is MBConv6, with k3×3, resolution of 28×28, 80 channels, and 3 layers. The simplified first backbone network model can greatly reduce the amount of calculation and make the output results fast and accurate.
[0050] 1.2) Extract 8, 16, and 32 times downsampled features from the first backbone network model constructed in step 1.1);
[0051] 1.3) The features extracted in step 1.2) are downsampled by 8, 16, and 32 times and input into the feature pyramid network for multi-scale feature fusion;
[0052] 1.4) Outputting the fusion result of step 1.3) to the detection head to detect and identify the target image;
[0053] The detection head uses a distributed Focal loss function. Considering that the true distribution is usually not too far from the annotated location, an additional loss function is added to enable the first backbone network to quickly focus on values near the annotated location, maximizing their detection probability. Based on this, the further optimized Focal loss function is the distributed Focal loss function.
[0054] The distributed Focal loss function reduces the number of shared weight parameters by 20%. However, since shared weights reduce detection capabilities, a set of convolutions is added to each layer of feature selection to improve feature extraction. The original Focal loss function uses group normalization in the detection head, which provides additional convergence benefits.
[0055] 2) Evaluate human posture;
[0056] 2.1) Build a lightweight network; the lightweight network consists of six layers of convolution:
[0057] The first layer is the Conv / Relu layer, with a kernel size of 3×3, a stride of 2, and 32 channels;
[0058] The second layer is the Conv / Relu layer, with a core size of 3×3, a stride of 1, and 64 channels;
[0059] The third layer is the Conv / Relu layer, with a core size of 3×3, a stride of 2, and 64 channels;
[0060] The fourth layer is the Conv / Relu layer, with a core size of 3×3, a stride of 1, and 64 channels;
[0061] The fifth layer is the TransConv / BN / Relu layer with a kernel size of 2×2, a stride of 2, and 64 channels.
[0062] The sixth layer is the TransConv / BN / Relu layer with a kernel size of 2×2, a stride of 2, and a channel of 32.
[0063] 2.2) Using a fast pose distillation model training method, the underlying pose is transferred from a pre-trained, larger teacher model to a constructed, lightweight network to form a student model. This fast pose distillation model training method uses heat maps of key human bone nodes for discrimination. Existing pose distillation methods are mostly based on category-level discrimination, while this method uses heat maps of key human bone nodes for discrimination. This approach makes pose distillation more efficient because it provides more feature constraints.
[0064] 2.3) Downsample the target image detected and identified in step 1.4 twice through strided convolution;
[0065] 2.4) Use the potential posture learned from the teacher model to train the student model; then press Figure 3 The process begins by feeding the target image, obtained by double downsampling in step 2.3), into the student model to obtain the confidence score of the target image. The student model then determines whether the input image pose meets the standard. The confidence score is then compared with a preset confidence threshold of 0.85±0.02. If the confidence score is greater than or equal to 0.85±0.02, an image display of the heat map of key bone nodes is generated, followed by a scoring system and a score output. If the confidence score is less than 0.85±0.02, a corresponding point deduction is performed, followed by a scoring system and a score output.
[0066] The teacher model is trained within the teacher network. The teacher network is an encoder-decoder network based on the real-time segmentation network model, used for motion feature extraction. The teacher local network is a component of the teacher network. The teacher local network is a stacked hourglass network, which is composed of two hourglass modules and obtains the residual of the original depth prediction. Each module of the lightweight network contains six convolutional layers and has a small receptive field of view.
Claims
1. A method for evaluating individual soldier tactical actions based on human posture estimation, characterized by: The following steps are involved: 1) Collect students’ body posture images in real time and identify target images; 1.1) Establish the first backbone network model; In step 1.1), the first backbone network model includes: The operator in the first stage is a 3×3 convolution operator, with a resolution of 224×224, 32 channels, and 1 layer. The operator in the second stage is MBConv1, with a core size of 3×3, a resolution of 112×112, 16 channels, and 1 layer; The operator in the third stage is MBConv6, with a core size of 3×3, a resolution of 112×112, 24 channels, and 2 layers; The operator in the fourth stage is MBConv6, with a core size of 5×5, a resolution of 56×56, 40 channels, and 2 layers; The operator in the fifth stage is MBConv6, with a core size of 3×3, a resolution of 28×28, 80 channels, and 3 layers; MBConv1 includes a 1×1 convolution, a batch normalization layer and an activation layer arranged in parallel, DWConv k×k, a batch normalization layer, an activation layer and a compression and excitation module arranged in parallel, a 1×1 convolution, a batch normalization layer and an addition operation module Add connected in sequence from the input end to the output end, and a side connection from the input end to the addition operation module; the step size of DWConv k×k is 1; the addition operation module Add is used to perform element-by-element addition of the input end and the output of each component from the input end to the output end to realize a residual connection; MBConv6 includes a 1×1 convolution, a batch normalization layer and an activation layer arranged in parallel, a DWConv k×k, a batch normalization layer, an activation layer and a compression and excitation module arranged in parallel, a 1×1 convolution and a batch normalization layer connected in sequence from the input end to the output end; the stride of DWConv k×k is 2; 1.2) Extract 8, 16, and 32 times downsampled features from the first backbone network model constructed in step 1.1); 1.3) The features extracted from step 1.2) with 8, 16, and 32 times downsampling are input into the feature pyramid network for feature fusion; 1.4) Outputting the fusion result of step 1.3) to the detection head to detect and identify the target image; In step 1.4), the detection head uses a distributed focal loss function to quickly focus on the value near the marked position to increase the probability of its detection; By adding a set of convolutions to each layer of features, it is used to compensate for the decline in detection ability caused by the distributed Focal loss function and improve the feature extraction effect; 2) Evaluate human posture; 2.1) Construct a lightweight network; the lightweight network consists of six convolutional layers: The first layer is the Conv / Relu layer, with a kernel size of 3×3, a stride of 2, and 32 channels; The second layer is the Conv / Relu layer, with a core size of 3×3, a stride of 1, and 64 channels; The third layer is the Conv / Relu layer, with a core size of 3×3, a stride of 2, and 64 channels; The fourth layer is the Conv / Relu layer, with a core size of 3×3, a stride of 1, and 64 channels; The fifth layer is the TransConv / BN / Relu layer with a kernel size of 2×2, a stride of 2, and 64 channels. The sixth layer is the TransConv / BN / Relu layer with a kernel size of 2×2, a stride of 2, and a channel of 32. 2.2) Transferring the underlying pose from a pre-trained teacher model to a constructed lightweight network to form a student model using a fast pose distillation model training method based on heat maps of key human bone nodes; 2.3) Downsample the target image detected and identified in step 1.4) twice through strided convolution; 2.4) The latent posture learned in the teacher model is used to train the student model, and the target image obtained by twice downsampling in step 2.3) is transmitted to the student model to obtain the confidence of the target image; the confidence of the target image is then compared with a preset confidence standard value. If it is greater than or equal to the confidence standard value, an image display of the heat map of the key bone nodes of the human body is output, and then the scoring system is entered and the score is output; if it is less than the confidence standard value, the corresponding point deduction operation is performed, and then the scoring system is entered and the score is output.
2. The method for evaluating individual soldier tactical actions based on human posture estimation according to claim 1, characterized in that: In step 2.4), the confidence standard value is 0.85±0.02.