A Lightweight Multi-Vehicle Detection Method Based on Multi-Scale Feature Fusion and Multi-Head Self-Attention

Aiming at the problems of high computational complexity, low recognition accuracy, and poor real-time performance of vehicle detection methods based on deep learning, this paper proposes a lightweight multi-vehicle detection method based on multi-scale feature fusion and multi-head self-attention. By adopting the FasterNet to replace the original CSPDarkNet53 backbone network of YOLOv8, and combining the idea of partial convolution, the C2f module in the neck is reconstructed to achieve efficient fusion of multi-scale features. After the last layer of the backbone network, a multi-head self-attention mechanism is introduced to enhance the ability of the model to understand complex spatial relationships and contextual information in images. Furthermore, the Focal-EIoU loss function is used to replace the CIoU loss function, accurately defining the aspect ratio of the predicted bounding box and alleviating the issue of imbalance between positive and negative samples. Experiment results show that the proposed method achieves 0.1% higher accuracy, 3.08% higher mean average precision, and 1.95 higher frames per second compared with the original model on the UA-DETRAC dataset.