Rep-YOLOv8: an enhanced object detector with deep feature representation for autonomous driving

Rep-YOLOv8: an enhanced object detector with deep feature representation for autonomous driving

연구 분야: Verification

학회: The Journal of Supercomputing

Accurate object detection is critical for the safety of autonomous driving systems in dynamic environments. However, in complex traffic environments, existing methods are susceptible to performance degradation because of their limited ability to express features. To address these challenges, we propose an enhanced detection framework called Rep-YOLOv8, based on the YOLOv8n architecture, which aims to systematically improve feature quality and detection performance. Specifically, we have developed a selective feature extraction module (SFE) to enhance the representation of features that are rich in semantic information while suppressing redundant interference. This approach alleviates the problem of semantic confusion during the multi-scale feature fusion process. To further improve the semantic modeling capability, we propose the Attention-driven Transformer-based Local Pyramid Fusion module (ATLPF), which combines the long-range dependency modeling capability of the transformer with the data-driven dynamic feature fusion strategy to achieve more context-aware feature interactions. In addition, a lightweight dynamic upsampling DySample is introduce to improve the spatial restoration capabilities of high-level features. Experimental results show that Rep-YOLOv8 improves mAP50 by 2.1% and mAP50:95 by 3.6% on the KITTI dataset. To assess its generalization, we further evaluated it on the BDD100K dataset, where it outperformed the original YOLOv8n. Furthermore, real-world driving tests also confirm that Rep-YOLOv8 strikes a balance between accuracy and efficiency, offering strong support for the industrial application of autonomous driving technologies.