LDTR: Transformer-based Lane Detection with Anchor-chain Representation

TL;DR

To enhance lane instance perception, LDTR incorporates a novel multi-referenced deformable attention module to distribute attention around the object and incorporates two line IoU algorithms to improve convergence efficiency and employs a Gaussian heatmap auxiliary branch to enhance model representation capability during training.

Abstract

Despite recent advances in lane detection methods, scenarios with limited- or no-visual-clue of lanes due to factors such as lighting conditions and occlusion remain challenging and crucial for automated driving. Moreover, current lane representations require complex post-processing and struggle with specific instances. Inspired by the DETR architecture, we propose LDTR, a transformer-based model to address these issues. Lanes are modeled with a novel anchor-chain, regarding a lane as a whole from the beginning, which enables LDTR to handle special lanes inherently. To enhance lane instance perception, LDTR incorporates a novel multi-referenced deformable attention module to distribute attention around the object. Additionally, LDTR incorporates two line IoU algorithms to improve convergence efficiency and employs a Gaussian heatmap auxiliary branch to enhance model representation capability during training. To evaluate lane detection models, we rely on Frechet distance, parameterized F1-score, and additional synthetic metrics. Experimental results demonstrate that LDTR achieves state-of-the-art performance on well-known datasets.

Figures (14)

• Figure 1: Lane prediction results using the current state-of-the-art method (CLRNet) and LDTR for cases with little- or no-visual-clue, lens flare, weak lighting, occlusion, and hidden lines. CLRNet misses certain lanes, while LDTR correctly finds all instances. Examples with ground truth are from the CULane dataset.
• Figure 2: Lane prediction results for CANet and LDTR for examples from the CurveLanes dataset. Limited by its lane representation, CANet cannot describe lanes in special cases like T-junctions, roundabouts, waiting areas, and sharp turns, while LDTR can address them all.
• Figure 3: LDTR follows the structural paradigm of DETR. After 2D image features are extracted by the backbone, LDTR further extracts deep semantic information in the encoder through the self-attention mechanism. The input object queries to the decoder are composed of content embeddings and anchor-chains. In the computation of each decoder layer, the object queries update themselves through MRDA and interact with image features, including the correction of anchor-chains and differentiation of positive or negative objects. After 6 iterative updates, the positive anchor-chains are able to represent lane shapes accurately. Additionally, LDTR introduces a Gaussian heatmap auxiliary branch to enhance the ability of the object query to perceive lane details.
• Figure 4: Various lane representations. It is hard for current methods to represent horizontal parts of lanes, but easy for our anchor-chain.
• Figure 5: The regression supervision approach of the anchor-chain enables it to efficiently utilize a small number of nodes to accurately describe curves, essentially similar to how humans recognize lanes.