基于视频异常理解的实验室不安全行为检测方法OA

Due to the presence of advanced equipment and chemical reagents in the chemistry laboratory,identifying unsafe behav-iors in the laboratory has become extremely complex.Accurately identifying and understanding unsafe behaviors in videos is a difficult problem for laboratory safety.Firstly,a dataset of laboratory behavior is constructed,and abnormal behaviors are annotated in detail based on VLM-Deepseek,including multiple-choice QA,anomaly types,temporal annotations,and inference chains.We use attention to ex-tract the positions of people and objects,and then performing spatiotemporal feature extraction.Three sets of such structures parallelized to improve extraction efficiency.Finally,we propose the understanding model of unsafe behavior based on GRPO,and different reward rules are formulated for reinforcement learning tasks includes multiple-choice QA,inference and classification.The model analyzes the cause of unsafe behaviors and gives reasonable explanations.The experimental results show that the accuracy of the spatiotemporal fea-ture extraction model is 96.1,and the classification accuracy based on the GRPO model is 75.21.It can perform more accurate behavior analysis and abnormal time localization,providing effective reference for intelligent management of chemical laboratory risks.