Human Motion Pose Analysis Based On Pose Estimation And Implementation On Edge Computing Platform

Computer vision technology has been applied in various domains of social life.Among these domains,the application of human pose estimation has garnered particular attention and has become an important research direction in computer vision engineering.For example,it finds applications in individualized sports assessment,behavioral rehabilitation monitoring and human-computer interaction.At present,an important focus of applied research is the design of low-complexity method for human motion analysis.In addition,considering personal privacy,human motion analysis based on infrared imaging has gradually become one of the research concerns.In this thesis,the Jetson Xavier nx edge computing platform is used as the computational infrastructure to delve into the study of lightweight and real-time deep neural networks.Specifically,we construct lightweight human motion analysis algorithms based on pose estimation in both visible light imaging and infrared imaging environments,aiming to meet the application demands in the aforementioned scenarios.The main contributions and innovations of this study are as follows:(1)We propose a method for human motion pose analysis based on pose estimation of deep learning.The method consists of two parts: pose estimation network and post-processing algorithm for human motion analysis.In the pose estimation network,we first design a dual-branch downsampling structure to increase the receptive field of network while reducing the floating-point operations,thus improving the detection efficiency.Secondly,we construct an enhanced feature extraction module after the main feature extraction backbone of the network.Its purpose is to integrate and enhance the extracted deep features of the human body,thereby improving the detection accuracy in complex environments.The design of post-processing method for human motion analysis combines human kinematics and fully considers the characteristics of different movements,and is responsible for mapping the positions of key skeletal points obtained from the pose estimation network to the limb states during human motion.Additionally,a threshold determination system is incorporated into the method to effectively avoid interference caused by human-shaped objects in the environment.Finally,the proposed method is deployed on mobile edge devices to accelerate the process of human motion analysis.The experimental results demonstrate the outstanding performance of the method in common human motion analysis tasks,with detection accuracy up to 90.9%,detection speed up to 77.6FPS,and remarkable anti-interference capabilities.Therefore,this method can accurately and efficiently accomplish tasks in practical environments.(2)We propose a human motion analysis method for thermal infrared imaging.Generalizing the method from visible light imaging to infrared imaging faces two main challenges,Firstly,the infrared imaging has echo phenomenon in the mirror environment,so that there will be multiple infrared reflections(artifacts for short in this thesis)which are very similar to the characteristics of the body in the infrared imaging.Aiming at this interference,we designe a two-stage lightweight neural network in this thesis to detect the position of infrared human artifact in advance,and utilizes image masking techniques to exclude these regions,The method is put into the front end of pose estimation network as the preprocessing stage.Secondly,infrared human body imaging is susceptible to false detections due to confusion with environmental objects.To solve this issue,a shallow feature extraction module with an attention mechanism is added to the pose estimation network,enhancing the accuracy of the detection results.Furthermore,network pruning is applied to retain the fusion parts of large-scale and medium-scale features,thereby improving the detection efficiency.Finally,the method is deployed on the edge computing platform,and its feasibility in infrared imaging is validated.Experimental results demonstrate that the method achieves a detection accuracy of 90.7%and a detection speed of 83.3 FPS,which meets the requirements of efficiency and accuracy for common human motion analysis in practical infrared imaging environment.