Motorcycle and Electric Bicycle Illegal Parking Detection System Using ROI-Based Object Recognition for Traffic Law Enforcement

Abstract

Illegal parking of motorcycles and electric bicycles on university campuses presents significant safety risks and logistical challenges. Conventional manual supervision remains inefficient, necessitating automated solutions. This study proposes an intelligent surveillance system leveraging a novel post-inference Region of Interest (ROI) strategy to detect violations in real-time. A custom dataset, constructed by combining three datasets (including Motor Itenas and Xe May), is used to train object detection models to detect vehicles and generate bounding boxes, without applying the ROI during training. The ROI module is applied after inference as a post-processing step to analyze detected objects and determine parking legality based on predefined zones. The methodology differentiates between vehicles in transit and those that are stationary by analyzing spatial relationships specifically overlap and angle between vehicles and pedestrians. Subsequently, it validates parked vehicles against predefined legal zones using dynamic aspect-ratio checks. A comparative analysis of YOLOv8L, Faster R-CNN, and SSD VGG16 architectures is conducted. Experimental results demonstrate that YOLOv8L delivers the optimal performance, achieving a superior F1-Score of 0.783 and mAP@50 of 0.815, while maintaining a viable real-time inference speed of 34.42 FPS. Conversely, while SSD VGG16 offers the highest speed at 57.68 FPS, it exhibited substantial accuracy loss. The findings validate that the proposed architecture agnostic approach provides a robust, scalable, and computationally efficient solution for automated traffic law enforcement, capable of operating effectively across diverse environmental conditions without requiring model-specific modifications.