An Unmanned Lane Detection Algorithm Using Deep Learning and Ordered Test Sets Strategy

Abstract

The traditional method of automatic lane detection is mostly based on Hough detection. However, this category of methods has low robustness and is vulnerable to interference. In order to improve the accuracy of lane detection, the presented paper compares and analyzes the end-to-end lane line detection network based on deep learning, including Unet-base and Deeplabv3+, in view of gradient explosion and slow running speed during model training, solutions are also given. Ordered test sets are used to speed up the training processing and validate the deep learning algorithm, in the case of different image resolutions, uses Unet-base and Deeplabv3+ to perform experiments respectively. Experiments show that under the same resolution, the Unet-base model with FCN network structure incorporating a better training strategy outperforms the Deeplabv3+ algorithm model that uses a classical ASSP module to solve the downsampling layer problem in terms of model generalization capability. And the MIOU of improved Unet-base is higher than Deeplabv3+. Therefore, compared to DeepLabv3+, the improved Unet-base model is more generalized.