Improved YOLOv8n-Based Multi-Target Detection Algorithm for Roads

Aiming at the problem of insufficient detection accuracy of multi-scale targets and long-range dense small targets in road scenes, this paper proposes an improved target detection algorithm based on YOLOv8n architecture. Firstly, Deformable Large Kernel Attention (DLKA) is invoked to focus on key target regions in complex traffic scenes through adaptive spatial weight allocation strategy to effectively suppress background noise interference; secondly, a new deformable convolution module (Gdyconv) is proposed to dynamically adjust the convolution kernel's sampling position and sensing field of the convolution kernel, which enhances the model's ability to capture geometric features of targets of different scales such as vehicles and pedestrians. Experiments based on the self-constructed road dataset show that the improved model improves the mAP50 metrics by 1.5% over the benchmark YOLOv8n, and the accuracy is increased by 1.2%, while the computational cost rises by only 3% with a 100% increase in the number of parameters. The algorithm achieves a good balance between accuracy and efficiency and provides a reliable solution for real-time multi-target detection in intelligent driving systems.