Overhead View Bus Passenger Detection and Counter using DeepSORT and Tiny-Yolo V4

Intelligent Surveillance Systems are increasingly gaining attention as countries build smart cities. The application of such systems includes monitoring foot traffic. This functionality enabled in a public transport setting allows better value-added service to passengers looking for bus availability with a guaranteed seat. With the computer vision technique applied, the innovation has room for an extension to other functionalities such as monitoring and predicting anomalies inside a public utility vehicle, whether it’s a security risk to passengers and drivers or a potential safety risk of a vehicle’s condition surrounding situations. This study presents the development of a people detection, tracking, and counting algorithm using a Tiny-YOLOv4, a deep learning object detection algorithm. The study aims to perform passenger counts through an overhead view of a camera. The results show that using Tiny-YOLOv4, a new custom model, as the base model proved the accuracy could be significantly improved, and false predictions can be eliminated or minimized in conjunction with the DeepSORT algorithm for tracking and counting tasks.