Research On Lane Detection Algorithm Based On Semantic Segmentation

In autonomous driving and advanced driving assistance system,lane detection is the key part of the lane departure warning and lane keeping assist modules,which has a significant impact on safe driving.In the past few years,deep neural network methods based on image semantic segmentation have shown the outstanding performance in lane detection task.In most of these network models,semantic feature in image can be learned and extracted via the heavyweight convolutional neural network(CNN),yet the efficiency and speed of detection will be affected using the heavyweight network model.Meanwhile,lane markings inherently have long continuous shape and this prior spatial information of lane markings is not fully exploited and fused using CNN.Furthermore,there is the certain correlation for lane markings position in continuous multi-frame images due to the driving scenes are continuous.In the CNN model,the information for time series in continuous multi-frame images is not fully considered and exploited.For the aforementioned problems,the research on lane detection network model based on semantic segmentation is made in this thesis,and the details are as follows:(1)The recurrent slice convolution module is proposed to fuse the spatial information,and the lane detection network model is constructed based on this proposed network module.The proposed recurrent slice convolution module uses the slice sequences from semantic feature map in horizontal and vertical direction as the input.The slice convolution unit in this network module fuses the adjacent slices from feature map and the spatial correlation information of lane markings in adjacent positions can be learned.The recurrent slice convolution module attains the lightweight level through the slice convolution with the shared weights.This proposed network module is embedded into the semantic segmentation network to establish the lane detection network model.In the detection network model,a distance loss function based on the spatial structure of lane markings is proposed.The proposed distance loss function can improve the stability and fitting performance of network training through the geometric constraints of lane markings.The experimental results show that the prediction accuracy on CULane dataset can be effectively improved by 6.2% via the proposed network module compared with the baseline model,and the running time of this module is 10 ms.Therefore,the proposed module greatly improves the detection performance and running speed of the lane detection network model.(2)The recurrent weighting convolution module is proposed to fuse the information for time series.The network module takes continuous multi-frame images as the cascaded input,and the weighting convolution unit in this network module is designed to weight and fuse the correlation information of lane position through the attention feature map of adjacent image sequences.In the recurrent weighting convolution module,the lane prediction network branches for each frame are designed to obtain the stronger attention feature of adjacent image sequences.An intermediate auxiliary loss function is used for the fully supervision in the training of these prediction network branches.The experimental results show that the performance of network is improved by 4.7% using multi-frame images compared with single frame image.(3)The combination mode of the recurrent slice convolution and recurrent weighting convolution network module is designed to construct the lane detection network model for the fusion of spatio-temporal information.Through temporal-spatial combination mode,the lane detection network fusing spatio-temporal information not only maintains the low complexity of the model,but also fuses and learns the spatiotemporal features of lane markings.The experimental results show that the effectiveness of the proposed spatio-temporal network model.Compared with the network model fusing the spatial information,the prediction accuracy on CULane dataset is improved by 2%.Compared with the network model fusing the temporal information,the prediction accuracy is improved by 1.4%.