Vision-Based People Counting System for Indoor and Outdoor Environments using different Deep Learning Models

In today’s world, identifying, tracking and counting people are essential aspects of video analysis and are highly sought after in the field of in Computer Vision. Occlusion is the major problem in people counting systems because it obstructs the visibility of individuals which leads to miscounting of individuals particularly in crowded environments. This study provides a comparison of three prominent deep learning architectures for people counting: Single Shot Detector (SSD), Faster Region-based Convolutional Neural Network (Faster RCNN) and You Only Look Once (YOLO) [different versions of YOLO; YOLOv3, YOLOv4, YOLOv5 and YOLOv8]. It comprises crowded and challenging scenarios to assess the strengths and weaknesses of each model in people counting tasks. The models are pre-trained. Firstly, we explored the SSD, a one-stage object detection model known for its simplicity. We have evaluated the accuracy and efficiency of SSD in people detection tasks. Secondly, we examined the Faster R-CNN model, a two-stage object detection approach renowned for its region proposal mechanism and the YOLOv3 model, a one stage object detection model. We analyzed its suitability for detecting people in crowded scenes. After comparing and examining these models, we found YOLO has the highest accuracy in people counting tasks. With the different versions of the YOLO model, we took numerous samples. The YOLOv8 produces significantly good results with precision 0.90, specificity 0.1 and $F 1$ score 0.93