Optimization of YOLOv8 Vehicle Detection Algorithm in Autonomous Driving Technology

In autonomous driving, accurate vehicle target perception matters a lot. This paper devises an enhanced vehicle target detection algorithm using multi-sensor fusion, seeking to address the high computational expense, weak generalization, and poor small target detection of traditional algorithms. Firstly, to bolster YOLOv8's global feature extraction, its backbone network was swapped for Swin Transformer. Then, a GSConv module paired with the channel attention SE replaced the regular convolutions in YOLOv8, trimming computational complexity and enhancing generalization. Finally, high-resolution feature layers were incorporated into the adjusted backbone network. These furnished higher spatial resolution and snagged more details. Multi-scale fusion of the added layers then upped small target detection. The improved YOLOv8 model's Precision, Recall, [email protected], and [email protected]:0.95 climbed by 3%, 5%, 3.5%, and 2.5% respectively, topping out at 88.3%, 86%, 89.9%, and 64.9%. Against traditional algorithms' drawbacks like high cost and weak detection, these gains were notable, validating the improved method.