Research And Application Of Lightweight Human Pose Estimation Algorithm

In recent years,convolutional neural networks have become the mainstream method of human body pose estimation,but the network structure is complex,and it has also brought a large number of parameters and calculations while improving the accuracy.It is difficult to support terminal devices with limited computing power and storage capabilities for real-time application.Therefore,from the perspective of model design,this thesis studies the lightweight and efficient human posture estimation algorithm based on the optimization of the balance between model calculation and accuracy,and implements an action similarity analysis system for fitness scenes based on this.The main work s of this thesis are of three folds.1.Considering the characteristics of the lightweight CNNs,we optimize Simple Baseline network.For the feature extraction module with huge parameters,we replace the backbone with lightweight networks.For the feature restoration module that consumes large computing resources,we perform channel compression on the transposed convolution layer with high computational complexity.After optimization,We finally obtain a lightweight and efficient network called Baseline-Lite.Our modified model only contains 3.8M parameters(1/9 of Simple Baseline)while the computational cost is only0.75GFLOPs(1/15 of Simple Baseline),which greatly reduces the computational overhead required for Simple Baseline at the expense of a small amount of accuracy.2.We explore the impact of different upsampling methods on the accuracy of keypoint localization.Then we propose PDFNet,a novel lightweight human pose estimation method based on proportional dense fusion.The innovation of this method lies in reducing redundant expression in multi-level feature fusion.Within one level,we adopt attention-aware method to enhance the discriminative capabilities of semantic and spatial features respectively.When performing multi-level feature fusion,we adjust the distribution of semantic and spatial features based on the scale ratio and then enhance the information exchange between multi-level features through pixelshuffle.Experimental results on COCO Keypoint Dataset indicate that the proposed method achieves 65.9m AP with only 0.43 GFLOPs computational cost,which is significantly ahead of other representative methods with the same lightweight backbone network.In addition,when expanding the backbone network slightly,this method can achieve 68.2m AP,which reaches the performance of mainstream methods with very few calculations and parameters.3.Based on PDFNet and the cosine DTW algorithm,we build a real-time action teaching system for fitness scenarios.The system can obtain the trainer's joint movement in real time,and analyze the similarity between the trainer's movement and the template based on the angle deflection between the joint vectors.The system provides scientific guidance for the learning of fitness movements and ensure the exercise standard of the trainer.