Research On Human Action Recognition Method Based On Bone Key Points

With the increasing emphasis on public safety,intelligent video surveillance systems have received more and more attention.As one of the important tasks of video understanding,human action recognition has become a research hotspot in the field of intelligent video surveillance.By identifying the actions that are happening within the monitoring range,it is possible to monitor and prevent emergencies in a timely manner,reduce the time cost of manual screening,and reduce the occurrence of sudden public safety incidents.However,due to the interference of background environment,mutual occlusion between human bodies,small targets and poor recognition of complex actions,the recognition effect is not ideal.In view of the above problems,this thesis studies the human action recognition method based on the data of skeletal modality,and proposes a human action recognition method based on skeletal key points.Since skeletal key point detection is an important part of action recognition based on skeletal points,this thesis improves the accuracy of action recognition from the perspectives of optimizing skeletal key point detection network and improving action recognition network.The main contents of this thesis are as follows :(1)Aiming at the problems of unbalanced model accuracy and efficiency and unsatisfactory detection effect of small targets in the existing bone key point detection methods,the HRNet high-resolution network is optimized.Firstly,the network is de-redundantly designed to reduce the complexity of the model when the detection effect is not affected.Secondly,a multi-scale feature supplement and blending module is added to make up for the accuracy loss caused by redundant design.Finally,a high-resolution feature pyramid structure is added to improve the detection accuracy of small target key points and ensure the detection effect of large targets.The experimental results show that the proposed DHRNet network has a recognition accuracy of 73.7% and an increase of 0.3% when the number of parameters and calculations only account for 41.8% and 76% of the original network.(2)Aiming at the problem that the existing action recognition methods based on skeletal key points have poor recognition effect due to weak time series modeling ability and mutual occlusion in a long time range,the Slow-Only network is improved.Firstly,the time convolution module is constructed into a hierarchical cascade structure,which increases the correlation of long time series information and improves the recognition accuracy of complex behaviors.Secondly,an improved channel attention mechanism module is added to make the network pay more attention to important features and reduce the interference caused by partial occlusion.The experimental results show that the improved network accuracy is increased by0.89%,reaching 94.3%.In addition,from the perspective of practical application,based on the above algorithm model,a set of human skeleton motion recognition system is designed,which proves the effectiveness and practicability of the proposed method.