Multispectral Road Segmentation with Spectral Attention Network

Road segmentation is a critical component of perception systems in autonomous driving, directly influencing the accuracy and safety of autonomous vehicle control. However, conventional segmentation methods based on RGB images face a significant challenge in distinguishing objects with similar color appearances but differing spectral characteristics, limiting their effectiveness in complex environments. To address these limitations, we propose a road segmentation framework that combines a multispectral imaging system with a spectral feature attention network. The framework exploits the rich material identification potential inherent in multispectral images and integrates a spectral attention mechanism to capture intrinsic spectral dependencies, which can enhance semantic representation. Specifically, it processes multispectral images with 9 spectral channels in the visible light range to enable fine-grained material discrimination. A spectral attention module adaptively assigns weights to different spectral channels, increasing the network’s sensitivity to critical spectral features and improving its performance in metameric scenes where traditional color-based cues are insufficient. To evaluate the proposed approach, we constructed a multispectral dataset for autonomous driving encompassing diverse complex environments. Comparative experiments with baseline RGB-based segmentation models demonstrate that our framework achieves improvements of 12.1% and 14.5% in Intersection over Union (IoU) and mean pixel accuracy (MPA), indicating robustness and material recognition capabilities. Overall, the proposed framework shows promise for advancing autonomous vehicle perception by effectively leveraging multispectral information and attention mechanisms to overcome the limitations of traditional RGB-based systems.