The accuracy and real-time performance of lane segmentation and lane type detection are still challenging subjects when the computing power of the autopilot computing platform is limited.To tackle these challenging problems, in this paper,we take each lane as a type of instance, transform the semantic segmentation task directly into an instance segmentation task, and propose a multi-task lane segmentation and detection network. The network can realize Three-branch tasks: lane instance segmentation, lane existence probability calculation, and lane type detection. In the network structure we designed, different branch tasks share the same feature extraction layer. The training phase synchronizes training in different branch serial modes to realize computing resource sharing and improve inference efficiency. We use a lightweight semantic segmentation network as our backbone network to reaching the balance of accuracy and computing, and it can be also worked when replaced and upgraded by other similar lightweight networks.We have evaluated our algorithm on TuSimple and Culane datasets and illustrated the comparison results with some existing state-of-the-art algorithms.Furthermore, we reported the trade-offs between accuracy and inferencing time of our network. In summary, we proposed a fast lane detection algorithm that can achieve high real-time performance and competitive accuracy. Keywords: Autonomous Vehicles, semantic segmentation, instance segmentation, lane detection, multi-task network

CICTP