Lightweight real-time lane detection algorithm based on ghost convolution and self batch normalization

摘要

A lane detection algorithm based on lane shape prediction with transformers (LSTR) is designed to address the problems of a large number of feature extraction network parameters, low utilization of original feature information, and easy loss of detail and edge information in the current lane detection algorithm. First, to reduce the number of parameters in the lane detection network and achieve a lightweight design, the ordinary convolution is replaced by ghost convolution (Ghost-Conv) with good performance; second, to enhance the utilization of the original feature information in the network and improve the lane detection accuracy, a self batch normalization (Self-BN) module is proposed to retain more original feature information by changing the normalization to achieve the improvement of the lane detection accuracy, and finally, to improve the accuracy of the network for lane detection, an efficient channel attention (ECA) mechanism is introduced to enhance the extraction effect of lane detail information and edge information. Experiments are conducted on the open source dataset TuSimple, and the results show that the proposed algorithm reduces the number of parameters and computation by nearly half, improves the detection speed by 33 FPS, increases the detection accuracy by 0.96%, reaches 97.11%, and reduces the false positive rate and the false negative rate by 0.55% and 0.71%, respectively, meeting the real-time requirements of autonomous driving, compared to the original network. Compared to other lane detection networks, it also has great advantages.