Abstract: To address the challenges of diverse and difficult-to-recognize merging and diverging locations in lane-level maps, this paper proposes a prior query-based detection network for merging and diverging locations.Firstly, candidate locations are recalled using geometric intersections of lane lines as anchors, and the surrounding lane information is rendered as images.Secondly, multi-scale image features are extracted using the ResNet network, and an encoder is employed to build long-range dependencies within the image features, thereby expanding the receptive field.Thirdly, a prior query encoding module is designed to extract the prior positional information of the candidate merging and diverging lane lines.Finally, the decoder is adapted to incorporate the prior positional information, and a binary classification task is designed to identify merging and diverging locations.Experimental results demonstrate that the proposed model exhibits strong noise robustness on maps with incomplete lane lines, achieving an identification accuracy of 96.27%.The proposed method demonstrates high accuracy and robustness in the identification of merging and diverging locations.

Key words: merging and diverging location detection, prior location, multi-scale, attention, end-to-end