Performance Analysis of YOLO Algorithms for Real-Time Crowd Counting

Abstract

Real-time head detection with counting based on crowded scenes is a very challenging and computationally complex task in the case of lengthy surveillance videos. Existing head detection methods suffer from slow detection and a high rate of missed detection, especially in the case of congested crowd regions as well as occluded heads. In this work, we performed performance analysis of various YOLO (You Look Only Once) architectures for real-time head detection and counting. We evaluated different YOLO architectures on standard datasets like SCUT_HEAD_A, SCUT_HEAD_B, and the Brainwash dataset. After experimental analysis, it is found that YOLOR outperforms by providing an mAP value of 0.91, 0.92, and 0.95 on the SCUT_HEAD_A dataset, SCUT_HEAD_B dataset, and Brainwash dataset, respectively..