A Multiscale Fusion YOLOV3-Based Model for Human Abnormal Behavior Detection in Special Scenarios

Abstract

Urban public security incidents are prone to occur. Better understanding of pedestrian abnormal behavior and trajectory in crowded places is conducive to crowd management and safety monitoring. A novel pedestrian abnormal behavior detection model (PABDM) is proposed to identify crowd behavior under abnormal scenarios. This model originated from a multiscale fusion you only look once (YOLO) version 3 (V3) algorithm and was trained using the PASCAL visual object classes (VOC) in combination with an abnormal pedestrian data set (APD), denoted as VOC+APD. Compared with YOLOV3-VOC, single-stage detectors (SSD)-VOC, and SSD-VOC+APD, the proposed model has notable advantages in prediction accuracy and detection efficiency. The results show that the network loss function of the model tends to be stable after 500 epochs, and its detection accuracy is 6% higher than the average accuracy of the compared models. This proposed model also effectively solves the problem of missing detection caused by edge target, fuzzy target, and small target in abnormal state human detection. The research results are of great significance for real-time crowd monitoring in complex scenes.