Research on Violence Detection Algorithm based on Multi-UAV

Abstract

Based on the 30-meter UAV aerial violence data, a multi-UAV violence detection algorithm based on target detection is proposed. In this paper, YOLOX algorithm is improved. Because the target pixel is small, the UFocus module is designed to upsample the feature image, and the channel attention module is cascaded after the UFocus. The DAM module is designed to improve spatial attention. The shallow features are extracted and cascaded with DAM modules, and multi-scale features are fused with the deep semantic features. A dual-threshold single-machine violence detection algorithm is proposed. When the number of violent actions detected in a violent video within three seconds meets the dual-threshold constraint, it is determined that violence has occurred. A violence detection algorithm based on multi-UAV information fusion is proposed. The confidence value of UAV is dynamically allocated, and the confidence and probability are fused to obtain the comprehensive results. The self-built violence detection data set is tested and trained. The experiment shows that the accuracy of the improved YOLOX algorithm is 1.56% higher than the original algorithm. The dual threshold detection violence detection algorithm can complete the detection task, with an average accuracy of 71%. Taking three UAVs as an example, the weighted fusion of multiple UAV information is carried out. When the probability threshold is 0.6, the accuracy rate reaches 89%, which is 25.4% higher than that of a single UAV.