Research On Key Techniques Of Soldier General Physical Training Evaluation Based On Deep Learning

Physical fitness training,as an important way for soldiers to improve their physical fitness,is the basic guarantee for the combat effectiveness of the army.In order to help the army better grasp the training situation of soldiers,it is necessary to use scientific and effective methods to evaluate training results.At present,the evaluation of soldiers’ physical fitness training results is mostly done manually,which has problems such as low efficiency,high consumption of human and material resources,and subjective factors affecting the evaluation results.This article applies deep learning technology to the evaluation process of soldiers’ physical fitness training results,targeting the general physical fitness training courses required for military officers and soldiers.The system collects soldiers’ physical fitness training videos through cameras,and applies pose estimation algorithms to obtain the coordinates of two-dimensional human body key points for pull-up,push ups,and sit-ups classes.Key frames of action videos are extracted for evaluation.For running courses,a multi-objective tracking algorithm is used to distinguish different soldier targets,track the soldier’s running process,and obtain the soldier’s running results.Based on the above research,a soldier general physical fitness training and evaluation system based on deep learning was designed.The main research content is as follows:(1)A Lite Trans Pose network model is proposed to address the issues of large parameter count,high computational complexity,complex training,and lack of inter frame connections in processing video sequences in Trans Pose networks.Lite Trans Pose uses lightweight network modules to reconstruct the convolutional backbone network portion to reduce network size while maintaining a high-resolution feature map processing structure to ensure model accuracy.A front layer normalization structure is adopted to reduce the hyperparameter sensitivity of Lite Trans Pose encoder in the warm up phase,which is more conducive to network gradient backpropagation and accelerates model convergence.At the same time,a multi head attention mechanism is introduced in the encoder structure to enable the network to focus on different subspace information and capture richer features.Finally,based on the optical flow propagation algorithm,the inter frame information of the video is connected to improve the data processing ability of Lite Trans Pose video streams.Through experiments,it has been proven that Lite Trans Pose network has a slight improvement in accuracy compared to Trans Pose network on COCO and MPII datasets,and the computational complexity has been reduced from 21.8G to 15 G,the parameter size has been reduced from 17.5M to 4.85 M,and the training resources have been reduced by 75%.The overall lightweight effect of the model has been achieved.(2)Aiming at the problems of low tracking accuracy and more fragmentation tracking times in DAN network,an ATT-DAN network model combining anti bottleneck structure and mixed attention mechanism is proposed.The ATT-DAN network uses the pre trained VGG16 as the feature extraction backbone network.In the process of feature dimensionality reduction,the anti bottleneck structure and channel attention mechanism are added.The anti bottleneck layer structure can reduce the loss of feature information.The channel attention mechanism helps the network adaptively learn potential key information and obtain more accurate tracking target appearance features.Introducing residual link structure in the compressed network to increase feature information utilization,while introducing spatial attention mechanism to reduce background interference,optimizing the weight of soldier target positions in the image,and improving network accuracy.Through experiments on MOT17 dataset,compared with the DAN network,the tracking accuracy of ATT-DAN network is improved by 8.0%,the number of tracking target ID switches is reduced by 81.6%,and the number of fragmentation tracks is reduced by 77.4%.ATT-DAN network achieves better tracking performance.(3)A universal physical fitness training and evaluation system for soldiers was designed based on the above network.Extract the human key point information of soldiers during pull-up and other physical training courses through Lite Trans Pose network,compare it with parameterized normative training action representation,determine whether the standard level of soldiers’ actions is within a reasonable range,determine whether the current action is standardized,and score accordingly.Track the position information of soldiers in running events using the ATT-DAN network,record the soldiers’ running trajectory,determine whether they have crossed the line based on the intersection detection results of the soldiers’ bounding boxes and runway lines,and record the soldiers’ running results.The soldier general physical fitness training evaluation system proposed in this article has good practicality through experiments in simulated soldier training scenarios.