Research On The Detection Method Of Non-suicidal Self-injurious Behavior Based On Indoor Activitie

Non-suicidal self-injury(NSSI)is a pressing global concern,and due to insufficient timely detection methods,patients often do not receive adequate attention and intervention.Research on NSSI detection using indoor behavioral patterns is valuable for psychology and computer vision fields.This paper examines self-injury behavior categories,establishes a dataset,and proposes an NSSI detection method using indoor spatiotemporal features.A lightweight network with improved depthwise separable convolution and a temporal shift module is designed,and an NSSI recognition algorithm is developed.Lastly,an NSSI recognition and early warning system is created.The main research findings include:NSSI detection algorithm based on indoor activity spatiotemporal features is proposed.To address the scarcity of NSSI datasets,a comprehensive NSSI dataset is collected and constructed.The algorithm is tested against various behavior detection methods and training strategies to determine the optimal combination.Results show that it achieves an m AP of97.98%.NSSI recognition algorithm based on a temporal shift lightweight network is proposed.Improved depthwise separable convolution,temporal self-attention,and temporal shift modules are combined to create the temporal shift lightweight behavior recognition network.Compared to four other behavior recognition models,the proposed algorithm achieves a Top1 value of 94.89% and a Top1 variance of 2.19 with only 32.14 M and 0.78 GFLOPs.Video inference tests show average CPU and GPU inference times of 127.37 ms and 9.82 ms,demonstrating robustness and recognition speed advantages.Based on the research findings,an NSSI detection prototype system is developed,integrating demand analysis,system architecture,and database design.Test results indicate that the system’s recognition accuracy reaches 84.18% and 88.33%.