The slender structure of lane makes it difficult to accurately locate in 3D space.Integrating high and low level features to capture global spatial structural relationships is beneficial for enhancing the accuracy of lane detection.To tack-le the above issues,3DLaneTR is proposed,which builds upon Anchor3DLane.3DLaneTR introduces a feature optimiza-tion module,enhancing the network's ability to extract global and detailed features.The anchor filter preserves the most representative and informative anchors,improving the prediction accuracy of the model.Furthermore,to fully utilize 3D spatial information,3DLaneTR introduces 3D position encoding through attention mechanism,enhancing the network's spatial perception ability.Results on OpenLane dataset show that F1 score of 3DLaneTR reaches 58.3％,which has in-creased by 9.8％compared to Anchor3DLane.
3D lane detectionheight estimationanchor filter3DLaneTR
NETL
NSTL
万方数据
