Research On Embedded AI Fitness Auxiliary Training System Based On Human Pose Estimation

With the development of our country’s economy and the improvement of people’s living standards,the public is increasingly concerned about health issues.Fitness is popular among the public as a common exercise to maintain health.For beginners,receiving effective guidance and feedback during the fitness process can not only help them master the movement,but also reduce the risk of injury during the process.Traditional fitness guidance relies on professional venues and professional coaches,and is not suitable for home environments,nor can it be carried out in the scenario of using fragmented time to exercise.At present,fitness software on the market can also provide fitness guidance,but most of them use the form of video playback,lack of effective supervision,and users cannot judge the accuracy and safety of their actions during exercise.Through the analysis of the research status at home and abroad,it can be seen that human pose estimation,as an important research content in the field of computer vision,is widely used in augmented reality to describe various movements and postures of the human body by finding the position of key points of a given object.,animation,games,and robotics.With the development of human pose estimation,many excellent network models have been born in the academic field,solving challenging problems such as different scales,crowded people,and ambient occlusion in key point detection.Although these network models have made outstanding contributions to solving the abovementioned problems,their model structure itself has the following problems:(1)The network models are consistently in-depth in the pursuit of accuracy improvement,and rarely consider network computing consumption and inference speed.,resulting in a more complex model structure,an increase in the number of parameters,and a cumbersome inference process.(2)The network model usually runs on a host or server with powerful computing resources and excellent performance.It does not care about the space occupied by the model,and the network mobility is not considered in the model design,which makes it impossible to run on resource-constrained mobile embedded devices.Therefore,aiming at engineering application,this thesis focuses on the balance between computational consumption and model accuracy.From the perspective of lightweight network structure design,the model parameters and model size are reduced while ensuring model accuracy,so as to ensure that the network model meets the deployment requirements.To the conditions of the operation of mobile embedded devices,in view of the above problems in fitness guidance,an embedded AI fitness auxiliary training system based on human posture estimation has been developed.The human posture estimation technology is used to supervise the user’s movement process through ordinary cameras,and real-time feedback is provided.Correct the user’s posture,and provide beginners with a good fitness experience anytime,anywhere in the form of twoway interaction.The main work and innovative achievements of this thesis are as follows:(1)From the perspective of lightweight network structure design,a lightweight human pose estimation network that balances computational consumption and model accuracy is obtained by optimizing Open Pose.This thesis innovatively designs two methods: multi-scale fusion link and DHSP Block.The backbone part in the original network is replaced by applying the multi-scale fusion connection method on Mobile Net V3,and the multi-stage part in the original Open Pose network is reconstructed by designing a DHSP Block.Finally,under the condition of ensuring the accuracy of the model,the amount of parameters is reduced by 78.4%,the amount of calculation is reduced by 93%,and the model takes up one-sixth of the original size,only 17.3MB.Possess the conditions for porting and deploying mobile embedded devices.(2)Quantize the model(model data type conversion)through Tensor Flow Lite,compress the model to 5.6MB with a slight loss of accuracy,and deploy it to the Android platform.Based on this model,an intelligent fitness auxiliary training system for mobile devices is developed.Through the angle-based heuristic algorithm,the functions of action recognition,real-time correction,and frequency statistics are realized for common fitness training actions.It runs stably at a frame rate of 30 FPS on Mi 10 Ultra with Snapdragon865.In addition,the system has designed a new AR interaction method(air gestures),which allows users to complete the training interface operation only by waving their limbs without touching the screen during the training process,providing a good user experience for fitness guidance in flexible scenarios.