A Novel Action Recognition Method Based On Bone Similarity Evaluation

The electrical substation is a key facility for power transmission and distribution.Unstandardized work behaviors of workers can lead to serious safety accidents.Action recognition is the basis of behavior judgment.Therefore,the research on automated human motion recognition is of great significance for the prevention of safety accidents.Currently,motion recognition methods face many challenges,such as insufficient data,diverse movements and postures,and high timeliness requirements.In response to these challenges,this article takes bone sequences as the research object,improves existing similarity assessment methods,and combines joint attention mechanism and relative entropy acceleration strategy to propose a motion recognition method based on bone similarity assessment.This method is applied to substation scenarios,providing strong support for the automatic identification of dangerous behaviors and safe production.The main work and innovations of this article are as follows:Firstly,aiming at the problem of excessive spatiotemporal overhead in classical similarity evaluation algorithms,a spatiotemporal optimization method for LDTW based on alternating multiplexing matrix is proposed.Using a dual channel distance matrix to reduce dimensions while alternately reusing to achieve the goal of requiring re storage and not storing,greatly reducing the time and space overhead of the Dynamic Time Warping Under Limited Warping Path Length(LDTW)algorithm while ensuring accuracy,laying the foundation for subsequent action recognition methods based on similarity evaluation.Secondly,to solve the alignment path problem,a method based on a dynamic pruning chain tree is proposed.In the LDTW algorithm,alignment paths need to be traced back and forward based on distance elements after the cumulative distance matrix is solved.This paper creatively uses a chain tree structure to represent alignment paths,and proposes a dynamic pruning strategy to reduce the storage space of the chain tree to efficiently solve alignment paths.Finally,Finally,a fast recognition method combining attention and relative entropy is proposed to address the issues of data diversity and high timeliness in motion recognition.Firstly,aiming at the problem of significant differences in spatial location and size of bone data,a method of constructing motion-independent invariant features is proposed,which aligns all bone data according to this feature to reduce the interference of motion-independent information on recognition;Secondly,a similarity evaluation model for bone sequences embedded in joint attention is proposed to enhance motion features and improve the robustness of motion recognition algorithms;Finally,based on a relative entropy acceleration strategy,invalid matching is minimized to ensure the efficiency of the algorithm.Experiments show that the accuracy of the method in this paper on the public dataset MSRC-12 and the proposed substation personnel common action dataset is90.32% and 95.42%,respectively,and the recognition time is 33.58 milliseconds and19.63 milliseconds,respectively.Its accuracy and time cost are superior to current mainstream benchmark methods,laying a solid foundation for subsequent research on behavior detection.