Research On Lane Line Detection Algorithm Based On Multi-task Learning

The automobile industry is an important pillar of the national economy.In the future,domestic automobile production will maintain a steady growth.The problem that the domestic automobile industry is big and weak is still outstanding.The development of intelligent vehicles is not only conducive to accelerating the transformation and upgrading of the automobile industry,but also conducive to enhancing the comprehensive strength of the country.In the multi-level assisted driving of intelligent vehicles,accurate lane line detection is the premise of preventing traffic accidents and reducing the accident rate,and at the same time,it can provide drivers with referential driving decisions.The lane lines in structured roads are taken as the object to promote the development of lane line detection algorithm,and the research is mainly carried out from three aspects of driveable road area segmentation,lane line detection,road area segmentation based on multi-task learning network and lane line detection.The specific work is as follows:(1)Design a path integral method based on the road the vanishing point,remove the redundant area in the road image with priority,and then inputs the image of the drivable road area into the neural network MultiNet,without any increase in deep learning network layer and change the training parameters on the basis of improve the learning ability of the network,the method also can be compatible with other network segmentation.Compared with other classic segmentation algorithms,this method improves the time it takes to run the entire segmentation network by nearly 60% while achieving good recognition accuracy.(2)Research on the improved lane line detection algorithm of Spatial CNN,and the blocked lane can be predicted by making full use of the lane structure of the network and the characteristics of the information transmission mode of the neural network.The improved network structure can realize the detection of multiple lane lines.Then the weighted least square method is chosen as the fitting method after discussing the common fitting method.Experimental results show that using F1 quantile as the evaluation index,the performance of Spatial CNN network is 0.6% higher than that of the original network,but it can support the detection of multiple lane lines.(3)Multitasking learning in transfer learning is introduced in this paper,combined with the previous discussion,a road area segmentation and lane line detection based on a multi-task learning network is proposed.The network is based on MultiNet as the main structure,and the improved lane line detection algorithm of Spatial CNN after adaptive modification is added.The convolutional layer with holes in improved Spatial CNN is replaced by ASPP structure.The experiment shows that the parallel multi-task learning network can promote the learning effect of branch network.For the lane line detection task,the F1 quantile as the evaluation index can improve the learning effect,which is about 0.9%,but it improves the robustness of each branch task.