Calculating Bus Occupancy by Deep Learning Algorithms

The increase in the world population causes unbalanced densities in public transport. While planning transportation networks, one of the most competing problems is to use an appropriate number of vehicles to reduce the increased load and not cause traffic. Although sensors or card reading systems are currently used to solve this problem, it will be more efficient to deal with real-life anomalies adaptively using Deep Learning algorithms. In this paper, we propose an approach consisting of three steps: person detection, tracking, and counting to automatically calculate bus occupancy. First, the person is detected by using the YOLO algorithm which is trained to detect persons in various environments. Second, the detected person is tracked by using the DeepSORT algorithm to determine if the passenger is leaving or entering the bus. Finally, the entrance and exit of passengers are counted by utilizing a virtual line drawn on the bus doors over camera images. The bus occupancy rate is estimated by comparing the current number of passengers with the maximum allowed number of passengers in every bus type. Two public datasets are used to test the proposed algorithm. These datasets contain images from embedded cameras that are fixed on public transport system buses. After testing the proposed algorithm on these two public datasets, our experimental results show that the proposed system can estimate bus occupancy with 93% and more accuracy.