Real-Time Automatic Detection of Motorcycle Helmet Based on Improved YOLOv8 Algorithm

In the field of road traffic, the application of a motorcycle helmet detection algorithm has a significant practical impact. This paper proposes an enhanced YOLOv8-ADown-LSCD algorithm based on YOLOv8 to address the shortcomings of existing detection algorithms, namely slow detection and the occurrence of false negatives. The algorithm utilises the ADown module in place of the convolution module in the traditional downsampling process, thereby reducing the calculation amount and the size of the detection target. This approach enables the achievement of the objective of improving the detection speed. Furthermore, a new detection head, designated as LSCD, is employed in place of the original detection head of YOLOv8. This replacement results in a significant reduction in the detection parameters, while maintaining the detection speed and accuracy. Comparative experiments have demonstrated that, when utilising the same data set, the average accuracy of the enhanced algorithm can reach 0.85 mAP@0.5. Furthermore, the detection efficiency and accuracy are superior to those of other algorithms, thus meeting the requirements of road traffic systems for motorcycle helmet wearing detection.