This paper addresses the problem of passenger detection and counting in urban public transport using in-vehicle surveillance video. Meeting this challenge would greatly optimize the efficiency of public transportation operations. We propose the Bus Passenger Counting System (BPCS) to solve this problem. The primary goal is to count passengers by employing computer vision techniques and deep learning model. The BPCS is specifically designed to handle scenarios characterized by significant occlusion, a common limitation in in-vehicle surveillance. To address this, multi-camera synchronization techniques were applied, enhancing detection accuracy through the complementary perspectives provided by overlapping fields of view. The proposed BPCS was evaluated on a real-world dataset captured from public buses. Results demonstrate that integrating multiple viewpoints and detection-based methodologies effectively meets the challenges such as occlusion and low-resolution imagery. The BPCS achieved a 94.1% accuracy rate in passenger counting which is enough for its practical applicability under real-world conditions. We conducted a comparison of YOLO-based architectures and found that the proposed model shows the highest performance, processing 5 frames in 0.62 s, utilizing only 1.1 GFlops on resource-limited platform Orange Pi 5 Plus.