A Multi-scale Enhanced Lightweight Lane Line Detection Algorithm

n order to reduce the imbalance between the real-time performance and detection accuracy of current semantic segmentation-based lane detection algorithms to adapt to engineering applications, a multi-scale enhanced lightweight lane detection algorithm (MEL-LDA) is adopted, and a lane detection method based on improved Res_Unet is proposed in this paper. First, a multi-scale enhanced atrous spatial pyramid pooling module (ME-ASPP), used at the encoding-decoding connection, is designed to improve the algorithm’s multi-scale feature learning ability. Secondly, we propose a model pruning strategy based on the global-local threshold, which can approximately reduce the model and solve the fault phenomenon caused by excessive pruning rate to achieve both reducing the algorithm’s complexity and maintaining the same accuracy. Compared with the Res_Unet algorithm, MEL-LDA improves the mean intersection over union (MIOU) by 1.3%, reduces parameter quantity by 61.1%, and improves single-image inference speed by 27.7%. Experimental results demonstrate that the proposed algorithm can improve the accuracy and inference speed of lane detection, effectively reduce the complexity of the algorithm model, and better meet the needs of practical production.