Addressing the issue of low recognition accuracy in violent behavior recognition algorithms for video data,this study proposes a method for violent behavior recognition based on multi-feature fusion of human nodes.The proposed method employs the YOLO-Pose algorithm to detect human bodies and estimate their poses,acquiring position information of human body nodes.It fur-ther extracts distance and shape features based on the human body structure and dynamic features as well as pose features based on motion characteristics from the nodes.All extracted feature information is subsequently fused.The behavior recognition model using Bi-LSTM is constructed to achieve the recognition and classification of violent behavior,while a behavior recognition result stabilizer is designed to address random interference-related misjudgments during recognition.To validate the effectiveness of the proposed ap-proach,experiments are conducted on the publicly available Violent-Flows dataset and a self-created dataset Vio-B for violent be-havior recognition.The results demonstrate that the proposed method achieves an accuracy rate of 97.9％and 98.5％on the Violent-Flows dataset and the Vio-B dataset,respectively,surpassing the performance of existing methods.
关键词
动作识别/暴力行为/特征融合/双向长短期记忆网络
Key words
action recognition/violent behavior/feature fusion/long and short term memory network
维普
万方数据