The lane detection algorithm based on multiscale aggregated attention fusion network

High accuracy and high stability are key elements of the lane detection algorithm in an autonomous driving system. Traditional algorithms are having difficulty extracting detailed features due to the complex geometric structure and background interference of lanes in real scenarios. Therefore, this paper proposes a Multiscale Aggregated Attention Fusion (MAAF) network, which integrates attention mechanisms to improve the accuracy and robustness of lane detection. Firstly, the Recurrent Feature-Shift Aggregator for Lane Detection (RESA) is improved to increase the effective sensory field and improve the efficiency of feature aggregation. Then, the ECANet attention module is used to extract features across channels, enhancing the model's focus on lane details. Finally, a spatial attention mechanism is incorporated to make the network more attentive to lane features, acquire more semantic information, and reduce the influence of background interference and clutter. Experimental results show that this method achieves 96.84% and 76.5% metrics on the TuSimple and CuLane datasets, respectively, surpassing the baseline network. Furthermore, it demonstrates good generalization and robustness, enabling accurate lane detection in complex road environments.