VBM-YOLO: an enhanced YOLO model with reduced information loss for vehicle body markers detection.

Abstract

In vehicle safety detection, the accurate identification of body markers on medium and large vehicles plays a critical role in ensuring safe road travel. To address the issues of the feature and gradient information loss in previous You Only Look Once (YOLO) series models, a novel Vehicle Body Markers YOLO (VBM-YOLO) model has been designed. Firstly, the model integrates the cross-spatial-channel attention (CSCA) mechanism proposed in this study. The CSCA uses cross-dimensional information to address interaction issues during the fusion of spatial and channel dimensions, significantly enhancing the model's representational capacity. Secondly, we propose a multi-scale selective feature pyramid network (MSSFPN). By a progressive fusion approach and multi-scale feature selection learning, MSSFPN alleviates the issues of feature loss and target layer information confusion caused by traditional top-down and bottom-up feature pyramids. Finally, an auxiliary gradient branch (AGB) is proposed. During training, AGB incorporates feature information from different target layers to help the current layer retain complete gradient information. Additionally, the AGB branch does not participate in model inference, thereby reducing additional overhead. Experimental results demonstrate that VBM-YOLO improves mean average precision (mAP) by 2.3% and 4.3% at intersection over union (IoU) thresholds of 0.5 and 0.5:0.95, respectively, compared to YOLOv8s on the vehicle body markers dataset. VBM-YOLO also achieves a better balance between accuracy and computational resources than other mainstream models, exhibiting good generalization performance on public datasets like PASCAL VOC and D-Fire.