Improved YOLOv8n Helmet Detection Algorithm
Changcong Zhang, Yan Hui Xu · 2024
Currently in complex scenarios, most existing helmet detection algorithms face challenges such as small target leakage, false detections, and low accuracy. In response, a novel helmet detection model named YOLOv8n-PSB is proposed, which improves upon the YOLOv8n model. To address the shortcomings of existing algorithms, several enhancements have been introduced in the YOLOv8n-PSB model. Firstly, a small target detection layer P2 has been incorporated into the backbone network to improve the learning of small target features by enhancing the fusion of shallow and deep semantic information. Additionally, the SPD-Conv module has been integrated to sample the feature map effectively without losing valuable information, thereby improving the detection of low resolution and small targets. Furthermore, the BiFPN bi-directional feature pyramid structure has been utilized to enhance multi-scale feature fusion, reduce background interference, and improve detection accuracy. Experimental results demonstrate that the YOLOv8n-PSB model achieves an average precision of 94.5%, with improvements in average recall, mAP0.5, and mAP0.5:0.95 by approximately 2% compared to the original YOLOv8n model. Overall, the enhanced YOLOv8n-PSB algorithm effectively enhances helmet detection in complex scenarios by reducing missed detections and false alarms for small targets, surpassing the performance of other common helmet detection algorithms.