Helmet Detection of Motobike Riders in Real-World Scenarios

Automated video surveillance systems for detecting helmet-wearing by motorbike riders and passengers play a critical role in enhancing road safety and increasing the efficiency of law enforcement. However, accurately detecting small objects, such as helmets, remains a technical challenge under varying environmental conditions. In this paper, we propose an ensemble strategy that combines a CNN-based model (YOLOvl0) with a transformer-based model (DE-DETRs) to detect violations of motorcycle helmet-wearing regulations. The DE-DETRs model utilizes multi-scale features to detect small objects, while YOLOvl0 offers real-time performance without the need for Non-Maximum Suppression (NMS). To combine the predictions from these two models, we apply the Weighted Box Fusion (WBF) technique, enabling more precise object detection. Additionally, to address the issue of data imbalance, we employ Curriculum of Data Augmentation (CUDA)-based data augmentation techniques to optimize the model's training performance. Experimental results demonstrate that the proposed ensemble model outperforms individual models, and WBF -based post-processing effectively improves detection accuracy across various scenarios.