Motorcyclist Helmet Violation Detection Framework by Leveraging Robust Ensemble and Augmentation Methods

Abstract

Traffic Monitoring Systems play a crucial role in real-life scenarios by improving traffic flow and reducing violations. Among these violations, helmet non-compliance is particularly common in countries where motorcycles are the primary mode of transportation. However, deploying an automatic violation capturing for helmet non-compliance in the real world presents challenges due to diverse traffic situations, environmental conditions, varying object sizes, and severely imbalanced datasets. In order to address these challenges, we propose a novel deep learning framework for helmet violation detection, which consists of four different modules, namely, object detection, object association, post-processing for tracking, and score correction. In particular, we develop a robust ensemble method to take advantage of various state-of-the-art object detection models such as YOLOv7, YOLOv8, Co-DETR, and EfficientDet. Furthermore, to address the issue of data imbalance, we propose two copy and paste data augmentation techniques for enriching data samples of rare classes. As a result, our approach yields a substantial 7.43% mAP enhancement over the baseline Co-DETR model, achieving a final score of 0.4792 in the 2024 AI City Challenge Track 5 test set and ranking 3rd among the competing teams.