Research And Application Of Human Action Recognition Technology Based On Deep Learning

Video-based human action recognition has always been a hot topic in the field of computer vision,which has been widely used in intelligent video surveillance,safe driving,human-computer interaction,video retrieval and content analysis and other fileds.With the development of deep learning,human action recognition technology has made breakthrough progress.But due to the complexity of human action and background noise,it is still a huge challenge on how to extract human action feature efficiently and design a human action recognition algorithm with high accuracy and generalization ability.In view of these problems,the main research work of this paper is as follows:A human action recognition algorithm based on 2D attention convolutional neural network is proposed.In order to reduce the interference of background and extract the key feature,this paper proposes a pixel-level attention block to drive the network autonomously learning feature weights.The algorithm effectively improves the performance of recogniton.A human action recognition algorithm based on 3D residual convolutional neural network is proposed.In order to make full use of the spatial and temporal information of video,this paper designs a 3D convolutional neural network,extracts spatiotemporal features,adds 3D residual block to improve performance and avoid vanishing gradients,and replaces the fully connected layer with global average pooling.The algorithm has a good effect and speed.A human action recognition algorithm based on multi-stream convolutional neural network is proposed.In order to enchance the generalization,this paper uses multi-modal input data,including RGB,optical flows and gradient maps,and uses 2D and 3D convolutional neural networks.This paper proposes two strategies on network weighted fusion.The accuracy on the UCF101 dataset and the HMDB dataset are 95.1%and 7 1.6%respectively.A warning mechanism for driver abnormal action based on sliding window is proposed.By adding up abnormal values and comparing with abnormal threshold,deciding whether to trigger warning.This paper designs and completes the prototype system,and complete the application-level verification of algorithm on the self-collected driver action dataset.