Lane Line Detection Algorithm Based on Image Binary Semantic Segmentation
A binary lane segmentation model combining ResNet34 encoding, feature aggregation, dual-path upsampling, and weighted cross-entropy improves OpenLane accuracy to 98.88% and MIoU to 66.97%.
Published Nov 8, 2024Paper ↗
Only vote on papers you've read. Sign in to vote.
Abstract
Lane line detection is susceptible to low detection accuracy due to illumination, extreme weather, complex traffic conditions, and imbalance in the number of lane line and non-lane pixel point samples. Therefore, this paper proposes a binary segmentation algorithm, EFAED, based on the ERFNet semantic segmentation algorithm with a coding-feature aggregation-enhancement-decoding structure, which consists of using the Resnet34 module to replace the ERFNet algorithm coding module, adding a feature aggregation module to its coding and decoding structure, replacing the original decoding module with a transposed convolution-interpolation upsampled dual-path feature map decoding module BU, and employing a weighted cross-entropy loss function. The experimental results show that compared with the original model, the image pixel accuracy is improved by 1.64% to 98.88% on the Openlane datasets, respectively, and by 3.21% to 66.97% on the MIoU, respectively.