Human Action Recognition and Custom Dataset for Bus Passenger Safety

Although many buses are equipped with onboard CCTV to assist drivers in monitoring the cabin, identifying abnormal passenger conditions in real time remains challenging. This work aims to enable early detection of emergency situations by recognizing passenger behaviors from video captured by cameras installed inside special-purpose vehicles. In practice, the interior of bus-like vehicles is highly complex and cluttered, which makes robust behavior understanding difficult. Another major limitation is the lack of publicly available datasets that represent diverse special-vehicle interiors. To address these issues, we recreated multiple bus environments and collected in-vehicle data to build a dedicated dataset. Using the collected dataset, we benchmarked several deep learning models to explore suitable approaches for passenger behavior recognition, and our proposed method achieved higher recognition accuracy than the competing baselines.