Tube-LaneNet: Predict each three-dimensional lane as a completed structure via geometric priors

Monocular three-dimensional lane detection is a critical task for intelligent vehicles. However, most current methods, which mainly extend the two-dimensional paradigms, regard lanes as separated points set and constrain loss through the orthogonal projection on the two-dimensional plane. In this work, a novel deep learning framework is proposed to detect each lane as a continuous completed three-dimensional spatial structure. Concretely, three-dimensional lane anchors are implemented to extract proposal features through geometric priors to guarantee the continuous linear spatial structure. To enhance the feature of proposals, a relation-aware mechanism is further introduced to extract the spatial relationship between three-dimensional lanes. In particular, a novel tube-like intersection over union (TubeIOU) is proposed, which extends each three-dimensional lane to the tube-like structure as a completed unified entity in the three-dimensional space. Experiments on different datasets demonstrate the state-of-art performance of the proposed framework, especially achieves the fastest efficiency with 69 frames per second. The code will be made publicly available.