Lane Detection and Tracking Based on Fully Convolutional Networks and Probabilistic Graphical Models

In this paper, we proposed a systematic algorithm that integrates a fully convolutional network (FCN) for lane detection and a probabilistic graphical model for lane tracking. The proposed method consists of three main steps: lane candidate extraction, multi-component lane modeling and multi-component lane tracking. The lane extraction step detects land candidates by fusing low-level lane edges extracted by traditional image processing and high-level lane segments extracted by the designed FCN. After lane candidates are detected and clustered, we divide each lane into n components and represent lane components plus whole lanes as nodes in the Hough transform domain. That is, we model each lane by a link structure with n+1 nodes. Finally, a probabilistic graphical model is built to track multiple lanes through frames in the Hough domain at the same time. The experimental results show that our method can track lanes in various challenging conditions in typical urban scenes such as curved lanes, texture marking, and lane occlusion. For comparison, we evaluated our system by Caltech Lane Datasets and show better performance in terms of metrics than previous techniques.