An Improved YOLOv8 Algorithm for Real-World Road Vehicle Object Detection

Vehicle detection plays a pivotal role in intelligent transportation systems and autonomous driving, ensuring safety, optimizing traffic flow, and supporting advanced driver assistance functions. However, dense vehicle distributions, occlusions, and differing target sizes within complex road environments often challenge the performance of existing detection models. In response, this work introduces an improved vehicle detection framework founded on YOLOv8s. Our design integrates a dual attention mechanism—merging CBAM and SE modules—into the Backbone, thereby reinforcing feature extraction and enhancing detection for smaller targets. Additionally, a cross-layer multi-scale feature fusion strategy, built upon ReP-GFPN, is incorporated into the Neck to boost multi-scale information sharing and strengthen detection across various target dimensions. We further replace the traditional CIOU loss with Wise-IoU, enabling the model to better handle difficult samples and occlusion scenarios. Experiments on the UA-DETRAC dataset demonstrate a 4.8% improvement in [email protected], underscoring the effectiveness of these enhancements in demanding traffic conditions. This work offers a promising avenue for advancing vehicle detection systems in intelligent transportation and autonomous driving contexts.