A MobileViT-Based Detection System for Motorcycle Traffic Violations

This paper proposes a real-time detection system for motorcycle traffic violations, which pose a significant threat to road safety. The system defines and detects three types of violations: signal violations, centerline crossing, and crosswalk violation. Road elements and motorcycles are detected using YOLO (You Only Look Once), followed by MobileViT (Mobile Vision Transformer)-based time-series analysis to interpret movement patterns over time. The system is built on ROS 2 (Robot Operating System 2) and operates in real time on embedded platforms such as Jetson Orin. Detection results are stored in JSON and CSV formats for further use. Experimental validation using actual blackbox driving footage demonstrated over 93% accuracy across various conditions. The integration of MobileViT effectively compensated for missed detections by capturing temporal patterns.