Research And Application Of Human Pose Estimation Algorithm In Fitness Scene

With the development of computer vision technology,the auxiliary fitness products with the function of motion teaching and motion evaluation gradually appear in people’s vision.The product can provide timely suggestions and feedback to fitness users on movements.The auxiliary fitness system mainly develops the application of fitness movement learning and fitness movement evaluation on the basis of human body posture.However,for the high dynamic information of human pose in the current fitness scene,the existing human pose estimation algorithm is still difficult to meet the actual needs of realtime and precision of the current fitness assistance system in terms of inference speed and stability.At the same time,the poor accuracy of the traditional similarity evaluation algorithm and the large amount of computation of the signal sequence similarity algorithm will also affect the real-time performance and stability of the auxiliary fitness system.Therefore,starting from the intelligent fitness assistance system,this paper improved and optimized the algorithms of human pose estimation,human motion similarity and motion sequence similarity,and realized the fitness assistance system with high precision and stability.The main contributions of this paper are as follows:(1)Aiming at the problem of slow inference speed caused by large number of parameters in human pose estimation algorithm based on high resolution network,this paper studied how to apply depth separable convolution and attention mechanism to high resolution network to achieve a large reduction in the number of model parameters while maintaining a slight decrease in accuracy.Then,according to the problem of low accuracy caused by occlusion of target nodes,a data enhancement mechanism of random occlusion was proposed to reduce the dependence of model on key point location features,so as to improve the accuracy of occluded nodes.The experimental results show that the reasoning speed of the lightweight model increases by 10.5fps,the number of parameters is 10.2%,and the prediction accuracy decreases slightly by 4.2%.The data enhancement strategy with random occlusion has certain advantages in the accuracy of the occluded key points.(2)On the basis of human pose estimation,pose similarity calculation and motion sequence similarity evaluation are studied.Firstly,to solve the problem that the traditional similarity assessment accuracy is not high and it is difficult to select features,a neural network-based model is studied to calculate human pose similarity.Secondly,for the problem that the calculation of dynamic time warping algorithm is too much,the adaptive adjusting region limit method is used to reduce the search region.Finally,experiments are carried out on UCF101 data set,and the experimental results show that the AUC index of the neural network method reaches 97.2% in the calculation of pose similarity.The experimental results show that the optimized dynamic time warping algorithm improves the inference speed by 1.4 times.(3)Preliminarily set up the fitness assistance system.Based on the improved algorithm of(1)and(2),a fitness movement learning and evaluation system is developed.The algorithm flow and system interface are introduced in detail.The experimental results show that the auxiliary fitness system can meet the requirements of accuracy and real-time fitness scene.