Performance analysis of human detection algorithms for OPDs in public Hospitals

Abstract

The Out-door Patients Departments (OPD) premises in public hospitals in Sri Lanka are frequently crowded with long queues of patients and the uncontrolled behavior of the crowd results to occur various difficulties. These issues can be easily managed with crowd detection techniques. Although multiple techniques are available for crowd detection, not all of them are applicable or effective in OPD premises. Therefore, the main goal of this work is to identify and evaluate the state-of-the-art algorithms for human detection in a crowd and counting, in order to determine the most effective algorithm for detecting and counting people at OPD premises. Three object detection algorithms YOLOv4, SSD Mobile Net v2, HOG-based SVM and sliding window were used for performance analysis. A dataset with 1000 images acquired from hospitals and vaccination centers covering a vast diversity of people with different positions such as sitting, standing, walking, laying, bending etc. were considered during image capture. The models were evaluated based on their mAP@0.50IoU score, total accuracy, and average recall. The experiments resulted that YOLOv4 outperformed the other models with a mAP@0.50 IoU score of 89.98%, total accuracy of 93.44%, and an average recall of 90%. The findings indicate that crowd detection in public hospital premises can be performed by the YOLOv4, which is a more effective object detection model compared to others.