Enhancing Traffic Law Enforcement through Deep Learning-Based Detection of One-Wheeler Motorcycle Stunts

Abstract

Abstract - Motorcycle-related road fatalities in Pakistan are alarmingly high, with reckless one-wheeling stunts being a significant contributing factor. Conventional traffic enforcement lacks the capability to detect, document, and act on such offenses in real time, creating an urgent need for automated solutions. This paper presents one of the earliest deep learning-based systems specifically designed for automated recognition and enforcement-ready documentation of one-wheeler motorcycle stunts. A custom dataset of 984 annotated images was curated from YouTube traffic footage captured across diverse lighting conditions, road environments, and geographic locations in Pakistan. The YOLO11m model was trained to classify two target categories, namely Normal Biker and One-Wheeler, and subsequently optimized for CPU-only deployment using the OpenVINO and ONNX frameworks. The system achieves a Mean Average Precision (mAP50) of 0.90 and mAP50-95 of 0.70, reflecting strong generalization under challenging, real-world traffic conditions. An automated evidence capture mechanism records up to five timestamped frames per detected violation, enabling actionable enforcement without manual oversight. On CPU-only hardware (Intel i3-8145U), the system operates at 1.2–5 FPS on recorded traffic feeds, establishing a viable and resource-efficient baseline for AI-driven traffic law enforcement in developing regions. The framework is readily extensible to additional violations such as helmet non-compliance, triple riding, and signal jumping, thereby contributing to a broader road safety ecosystem.