YOLO-VSF: An Improved YOLO Model by Incorporating Attention Mechanism for Object Detection in Traffic Scenes

Intelligent transportation and autonomous driving systems have made urgent demands on the techniques with high performance on object detection in traffic scenes. This paper proposes an improved object detection model YOLO-VSF over the YOLOv4 model, which is a representative work with excellent performance among YOLO series of object detection models. The main improvement measures include: The backbone feature extraction network CSPDarknet53 of YOLOv4 is replaced with VGG16 to improve the feature extraction capability; SENet attention mechanism is incorporated to improve the salient and correlation feature representation capability; Focal Loss is integrated into the loss function to overcome the sample imbalance problem. In addition, the detection performance of small targets is improved by increasing the resolution of input images. Experimental results show that on the VanJee traffic image dataset provided by Beijing VanJee Technology Co., Ltd., the proposed YOLO-VSF model achieves an average mean accuracy (mAP) of 92.21 percentage points, which improves the mAP by 3.04 percentage points compared with the YOLOv4 model while maintaining the detection speed of the original model. On the UA-DETRAC dataset, the average accuracy of YOLO-VSF is close to that of the latest YOLOv7 model with the number of parameters reduced by 1.329 × 10⁷. The proposed method can provide a support for object detection in traffic scenes.