Intelligent Crowd Analysis: A Learning-Based Approach for Detection and Behavior Analysis in Surveillance Videos
Crowd monitoring and analysis is a recent initiative that brings public safety as a result (especially if your business is in a crowded area) This work present a novel Intelligent Crowd Analysis (ICA) system in this paper that takes advantage of a Modified YOLOv4-tiny object detection model with Non-Maximum Suppression (NMS), Deep SORT object tracking and a Crowd Monitoring and Behavior Analysis (CMBA)module that is designed by us. The system enables the detection of real-time violations in social distancing, entry in restricted areas, and abnormal crowd behavior. In order to tackle these challenges, such as overlapping objects, dense crowds, and dynamic conditions, some domain-specific changes, such as setting anchor boxes manually and adding attention to the split detector networks, have been made. Experiments show that the system is effective and robust in different surveillance situations. When comparing the results of the Modified YOLOv4 with the state-of-the-art models like YOLOv3, Faster R-CNN, SSD, and FairMOT, the Modified YOLOv4 achieved the best precision (96.87%), recall (95.31%), F1-score (96.08%), and accuracy (97.46%) within no time while the real-time video/frame or image processing was performed through deep learning on individual vehicle counting providing better performance of results over these methods [153]. The system runs with an average of 25 FPS; thus, real-time capabilities are guaranteed. In addition to these totals, the performance achieved a 92% detection accuracy, identified situations of social distance violations, Restricted Area Violations, and abnormal motion with highly Reliant Identification. This proves that this intelligent surveillance system operates in very complex environments and in real-time, which makes it suitable for public safety, smart cities, and event management-related applications. The ICA system demonstrates improved performance in crowd analysis and monitoring.