An AI-Driven Edge System for Improving Safety in Public Bus Transport

Abstract

Ensuring passenger safety is a key requirement for sustainable public transportation; however, the problem of detecting violent behavior in buses has received limited attention. Most existing methods are developed on generic CCTV datasets and rely on computationally demanding models, often neglecting the real-time constraints of embedded systems used in transit scenarios. In this work, we propose an edge-based artificial intelligence system for real-time violence detection tailored to urban buses. The system integrates a multi-camera configuration with onboard processing, removing the need for cloud-based computation. To address domain shift, several state-of-the-art 3D convolutional neural networks are trained using a combination of public datasets and proprietary data collected within a full-scale bus simulator. Experimental evaluation, including field tests on a 13-meter bus, demonstrates that incorporating transport-specific data significantly enhances model generalization. Among the tested architectures, X3D achieves the most favorable tradeoff between accuracy and computational efficiency, enabling sub-second inference on embedded platforms.