HFCA-LSTR: an infrared lane detection method integrating high-frequency enhancement and attention mechanisms

Abstract

To address the challenges of low lane-line contrast, blurred edges, and strong interference from background thermal noise in infrared scenes, this paper proposes an infrared lane detection method termed HFCA-LSTR, which integrates high-frequency enhancement and attention optimization. Built upon the end-to-end lane detection model LSTR, the proposed method improves the network from two aspects: shallow edge feature enhancement and deep semantic feature refinement. First, a High-Frequency Cross Attention (HFCA) module is introduced into the shallow layers of the backbone network. By exploiting the Fast Fourier Transform (FFT) and a learnable high-pass filter, the module extracts high-frequency edge information and enhances the local texture representation of infrared lane lines through spatial-channel interaction. Second, a Convolutional Block Attention Module (CBAM) is embedded into the Transformer encoder to recalibrate encoded features along both channel and spatial dimensions, thereby reducing the interference of thermal background noise in lane representation. In addition, an infrared lane dataset named Infrared is constructed for model training and evaluation. Experimental results show that HFCA-LSTR achieves a detection accuracy of 96.04% on the Infrared dataset, improving the baseline LSTR by 1.89 percentage points. Furthermore, when visible-light pretraining and infrared full finetuning are adopted, the accuracy further increases to 96.44%, while maintaining a high inference speed of 360 FPS. These results demonstrate that the proposed structural improvements effectively enhance infrared lane detection performance, and that full fine-tuning can further boost the model on this basis.