An Improved Dual-Subnet Lane Line Detection Model with a Channel Attention Mechanism for Complex Environments

Recently, many lane line detection methods have been proposed in the field of unmanned driving, and these methods have obtained good results in common conditions, such as sunny and cloudy conditions. However, these methods generally perform poorly in poor visibility conditions, such as foggy and rainy conditions. To effectively solve the problem of lane line detection in a foggy environment, this paper proposes a dual-subnet model that combines a defogging model and a lane line detection model based on stacked hourglass model blocks. To strengthen the features of important channels and weaken the features of nonimportant channels, a channel attention mechanism is introduced into the dual-subnet model. The network uses dilated convolution (DC) to reduce the network complexity and adds a residual block to the defogging subnet to improve the defogging effect and ensure detection accuracy. By loading the pretrained weights of the fog-removing subnets into the dual-subnet model, the visibility is enhanced and the detection accuracy is improved in the foggy environment. In terms of datasets, since there is currently no public dataset of lane lines in foggy environments, this paper uses a standard optical model to synthesize fog and adds a new class of foggy lane line data to TuSimple and CULane. Our model achieves good performance on the new datasets.