An Improved Lane Detection Algorithm Based on Hybrid Model

To minimize the incidence of erroneous detections of lane identification algorithm under different road environment such as shadow occlusion of lane line and white phenomenon of road surface. This paper presents an improved hybrid deep neural network model for lane identification by utilizing a series of consecutive frames captured from driving scenes. Based on SegNet-ConvLSTM network, this paper introduced Focal loss function and added improved channel attention mechanism and Bp-ASPP pyramid, so that the enhanced network could further extract image features and aggregate context information of different regions on the basis of the original network. Experiments were carried out on TuSimple dataset, and the accuracy of the proposed method was significantly improved compared with the previous method, which verified the proficiency of the approach.