Authors: Assistant Professor Dr.B.Nageshwar Rao, E. Prathyusha
Abstract: Vision-based people counting has become an essential component of intelligent surveillance systems due to its wide range of applications in public safety, crowd management, smart cities, transportation, retail analytics, and access control. Accurate counting of individuals in both indoor and outdoor environments remains a challenging task because of factors such as occlusion, varying illumination, background clutter, and changes in camera viewpoints. Traditional image processing techniques often fail to provide reliable performance in crowded scenes and dynamic environments. This study presents a deep learning-based vision system for people detection and counting by comparatively evaluating Single Shot Detector (SSD), Faster Region-based Convolutional Neural Network (Faster R-CNN), and multiple versions of You Only Look Once (YOLO), including YOLOv3, YOLOv4, YOLOv5, and YOLOv8. The proposed framework consists of data acquisition, image preprocessing, person detection, centroid tracking, and people counting modules. Experimental analysis is conducted using indoor and outdoor image and video datasets collected under diverse environmental conditions. The comparative evaluation demonstrates that YOLOv8 achieves the best overall performance with an accuracy of 0.92, precision of 0.90, and F1-score of 0.93, outperforming SSD, Faster R-CNN, YOLOv3, YOLOv4, and YOLOv5 in terms of detection accuracy and counting reliability. The proposed framework provides an efficient and scalable solution for real-time people counting while preserving the original methodology, deep learning models, datasets, and experimental findings presented in this study.
