基于B-RESA的车道线检测模型OA

Lane detection is a fundamental task in autonomous driving and intelligent transportation systems,requiring both high accuracy and real-time performance.Despite the widespread adoption of deep learning techniques,achieving an optimal trade-off between detection accuracy and inference efficiency remains a significant challenge.A novel lane detection framework based on the bidirectional recurrent feature-shift aggregator(B-RESA)is proposed,inspired by the RESA architecture.Specifically,B-RESA performs bidirectional recurrent feature shifting along both vertical and horizontal axes of the feature maps,enabling the network to capture long-range spatial dependencies more effectively.To further improve feature aggregation,a directional-aware weighting(DW)module that adaptively fuses features from different directions based on learned importance weights is introduced,thereby enhancing the model's robustness in complex road conditions.Extensive experiments on the CULane and TuSimple datasets demonstrate that the proposed method achieves superior performance in both detection accuracy and inference speed compared to existing approaches.Ablation studies confirm that the DW module contributes significantly to performance gains,and that B-RESA notably accelerates inference without sacrificing accuracy.