Lane Detection Algorithm Based on UNet-CBAM-ASPP Network

Aiming at the problems of incomplete feature extraction and insufficient context-associated information in lane line detection, this paper improves on the basis of UNet and proposes a lane line detection algorithm UNet-CBAM-ASPP based on multi-scale feature fusion and attention mechanism. The algorithm adopts the CBAM attention mechanism, which is capable of dynamically adjusting the importance of each channel and spatial location in the feature map, and is able to effectively suppress interference, while enhancing the ability to pay attention to and extract subtle features. In addition, the algorithm also introduces the multi-scale feature fusion module ASPP, which expands the sensory field by using cavity convolution, focuses on fusing different scale features to enhance the expression of semantic information, and adopts the Dice Loss functions and the Focal Loss functions to prevent sample imbalance. Finally, simulation experiments are performed using the CULane and the CurveLanes datasets. The simulation results show that the network in this paper detects lane lines with 98.21 % accuracy and 80.12 % precision, which is much higher than the mainstream segmentation models, such as DeeplabV3+, PSPNet and HRNetv2.