Reb-YOLO: An Accuracy-Improved Vehicle Detection Network Under Low-Recognized Environments

The vehicle detection under low-recognized environments is a great challenge for the reliability and effectiveness of advanced driver assistance systems (ADASs). To address this issue, this article establishes a reparameterized and bidirectional (Reb) feature fusion you only look once (YOLO) network to enhance vehicle detection performance under low-quality background. This novel network incorporates three components: a simple attention module in backbone (Sim-backbone), a reparameterization method-integrated neck (Rep-neck), and a lightweight decoupled head. First, to reduce interference under low-quality environments, we apply 3-D weights to the feature maps, and a Sim-backbone is created. Second, a cross-level connection and multibranch method are deployed to enhance feature fusion and feature representation, improving the network’s perception of small and incomplete vehicle targets, with the development of Rep-neck. Third, GhostConv and depthwise separable convolution (DWConv) are, respectively, deployed in localization and classification branches of the decoupled head, this ensures more accurate target localization and classification while achieving lightweight design and mitigating overfitting issues associated with increased model complexity, and a lightweight decoupled head is established. To assess the performance of the proposed network, experiments were conducted by utilizing the curated BDD100K dataset, KITTI dataset, and VOC2007 dataset. The results demonstrate that the proposed network outperforms the baseline network YOLOv8n and other advanced networks in detecting vehicles under low-recognized environments, and the proposed network is more lightweight.