A Coarse-to-fine Two-stage Helmet Detection Method for Motorcyclists

Hongpu Zhang, Zhe Cui, Fei Su; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 7066-7074

In recent years motorcycle accidents have occurred frequently with a important reason being that motorcyclists do not wear helmets properly. The visual method of detecting whether a motorcyclist is wearing helmet based on monitoring videos can provide technical support for traffic management. However the appearance characteristics of motorcycle drivers and passengers are too similar to distinguish which makes it difficult to detect helmet. In this task we propose a Coarse-to-fine Two-stage Helmet Detection Method for Motorcyclists to improve the accuracy of helmet and motorcyclist detection. Our Coarse detector detect the rough location of people and motorcycle as the initial suggestion for the following Fine-grained detection. Then our Fine-grained detector employs a classification branch to accurately distinguish between the driver and passengers. Finally we use some useful strategies such as Test Time Augmentation (TTA) and Weighted Boxes Fusion (WBF) to achieve further improvements to our proposed framework. Our proposed framework achieved mAP score of 39.4% on the test dataset of AI City Challenge 2024 Track5.