Small Object Vehicle Detection Based on Improved YOLOv5

With the advent of deep learning, object detection has been extensively utilized in intelligent transportation systems. However, the performance of baseline object detection models in traffic scenes is suboptimal, particularly for small object detection. To address this issue, this paper introduces an improved method for small object detection in traffic scenes based on YOLOv5, designated as YOLOv5_CBS. Initially, the convolutional modules of the model network are replaced to enhance the capability of feature extraction. Subsequently, an attention mechanism is incorporated into the Neck section of the model network, which bolsters the network's ability to capture contextual information through bidirectional encoding and decoding. Additionally, the imbalance between simple and challenging samples is mitigated by introducing a weighting function, Slide, into the model's loss function. Experimental results on the SODA10M dataset demonstrate that YOLOv5_CBS achieves average detection accuracies of 64.3% for mAP@0.5, 46.3% for mAP@0.5:0.95, and 16.2% for small object detection, representing improvements of 4.3%, 7.8%, and 2.1% over YOLOv5, respectively.