Research On Video-based Action Recognition Method For Edge Environment

In recent years,video action recognition technology has undergone rapid development and has been widely used in many fields such as security,medical treatment,and human-computer interaction.In the past,action recognition services were mostly deployed in cloud centers.However,as the research of action recognition gradually shifts from offline processing to real-time calculation and analysis of video streams,and a large number of video data sources are gradually transferred to edge nodes,cloud computing is becoming unsuitable for action recognition services.Because cloud services far away from data sources may face network congestion and insufficient bandwidth when receiving data,they cannot meet high real-time requirements.Cloud services far away from the data source may face network congestion and insufficient bandwidth when receiving data,so that they cannot meet the high real-time requirements.Edge computing is an emerging paradigm that transfers computing tasks and services from cloud centers to edge nodes.It has the advantages of low latency and high service security.Using edge computing to deploy behavior recognition services on edge devices just makes up for the deficiencies of cloud computing.However,most of the existing action recognition methods are based on deep learning and rely on a large number of parameters and calculations.The deployment of action recognition services on edge devices has to consider the problem of edge device resource constraints.Therefore,it is of great significance to study lightweight action recognition methods.Aiming at the problem of limited resources in the edge environment,this paper focuses on the research of lightweight video action recognition methods.According to the direction of this subject,the main research contents of this article are as follows:(1)Video action recognition method based on DSconv-GRU is proposed.The method mainly designs a GRU structure(DSconv-GRU)that combines self-attention mechanism and depth separable convolution,and optimizes the data processing flow of the model.That reduces the demand for edge device resources from the video action recognition service from the perspective of the lightweight model.Experimental results show that this model achieves a better lightweight effect.It enables the action recognition service to have better recognition accuracy and a smaller resource occupancy rate.(2)A Sliding Window with elastic jumping for Online Detecting Action Start is proposed.This method controls the input of the video stream by allowing the sliding window to adopt a elastic jumping mechanism.In this way,the Online Detecting Action Start service can reduce the detection of redundant data by the model without losing key data,thereby reducing the service's consumption of edge device resources.Experimental results show that this method reduces the service's occupation of computing and storage resources while keeping the action recognition performance similar to popular methods.