Action Recognition And Evaluation Based On Deep Learning

The rapid development of deep learning has provided an opportunity for the implementation of AI services,and the popularity of mobile devices has brought about an exponential increase in traffic.As the biggest traffic booster,video has become a research hotspot in the field of AI.Behavior understanding,video summarization,video scene classification,subtitle generation and other technologies have been widely used in smart medical care,autonomous driving,virtual reality and other fields.While bringing huge traffic,it also brings new opportunities for video-based AI tasks.needs and challenges.This paper adopts deep learning technology,based on human behavior recognition and evaluation tasks,to study the recognition accuracy and efficiency under multi-source data to meet the diverse needs of different application scenarios,including cloud with sufficient computing power and services with limited resources end and so on.In addition to theoretical research,this paper also designs a youth-oriented sports intelligent service system that integrates automation and standardization,and realizes the implementation of the algorithm.The specific work is as follows:1.An adaptive scale graph convolution behavior recognition algorithm is proposed.Cloud computing power can support large-scale graph computing and deep networks.Therefore,this algorithm aims to maximize the accuracy of behavior recognition by improving the network structure and provide high-precision behavior recognition services for cloud devices.Aiming at the lack of flexibility of graph scale in graph convolution action recognition algorithm,an adaptive scale graph convolution module and a multi-scale fusion module are proposed.The former is based on a priori and spatial attention mechanism to construct an activity judger to adaptively change the scale of the skeleton map;the latter uses a multi-channel attention mechanism to dynamically fuse multi-scale features.In addition,multi-channel features such as keypoints,bones,and motion information are also extracted to enrich the input of the network.The algorithm has achieved obvious improvement on the NTU-RGBD dataset.2.A lightweight human behavior recognition algorithm based on genetic algorithm is proposed.In order to meet the needs of low-latency applications,this algorithm designs a behavior recognition model with low complexity,small computational load,and accuracy that meets the requirements of use on the resource-constrained server side.Based on network architecture search and genetic algorithm to achieve lightweight constraints on the model.Aiming at the problem of high resource consumption and long time-consuming in the process of network architecture search,an improved genetic algorithm based on online agent is proposed,which uses the agent model to pre-screen disadvantaged individuals and reduce the number of candidates.According to the lightweight requirements of the model,a multi-objective genetic algorithm is designed to control the amount of model parameters,and the effectiveness of the algorithm is verified on the Opportunity dataset and UniMiB SHAR dataset.3.A sports intelligence service system for teenagers is designed.In view of the lack of unified standards and low automation rate in the existing physical fitness test and evaluation,we propose to use computer vision technology to automatically monitor and evaluate students’ physical fitness,posture,body shape,etc.Specifically,we use person re-identification,pose estimation and behavior recognition to monitor and test the specific actions,which can not only reduce labor costs but also improve the efficiency of physical testing for primary and middle school students.