YOLOv8-FDD: A Real-Time Vehicle Detection Method Based on Improved YOLOv8

Abstract

Aiming at the serious problems of missed detection, false detection and difficult deployment of existing target detection algorithms when applied to traffic scenes, a vehicle detection model YOLOv8-FDD with lower parameter count and higher accuracy is proposed in this paper. First, an integrated Feature Sharing Detection Head is proposed to compress the amount of redundant parameters in the model head while ensuring the accuracy of the algorithm. Second, to enhance the interaction between regression features and classification features in the detection head, an Feature Dynamic Interaction Detection Head is further designed based on the Feature Sharing Detection Head. Third, the Dilation-wise Residual (DWR) module to strengthen the multi-scale feature extraction capability of the backbone network. Finally, the dynamic upsampling module DySample is utilized to improve the nearest neighbor interpolation upsampling in YOLOv8. Validation on the UA-DETRAC dataset shows that the proposed YOLOv8-FDD model’s parameter count is only 72.89% of the original YOLOv8, with computational complexity slightly increasing from 8.1 GFLOPs to 8.5 GFLOPs. The mAP50 and mAP50-95 improved by 0.7% and 1.3% respectively, with FPS consistently above 300. Additionally, on a self-built traffic surveillance dataset, YOLOv8-FDD outperformed the original algorithm. The error type statistics indicate that YOLOv8-FDD significantly reduces both vehicle false negatives and false positives, making it a high-accuracy, well-generalized real-time vehicle detection method.

Keywords

• Deep learning
• vehicle detection
• YOLOv8
• DWRSeg
• DySample