LiDAR-camera fusion based high-resolution network for efficient road segmentation

This paper addressed the problem of road segmentation using a novel LiDAR-Camera fusion based high-resolution network. Road segmentation in different road conditions has been challenging due to limitations of single sensor. LiDAR could detect height and distance accurately in all road conditions but its data is too sparse for segmenting road, and camera can capture rich visual features but is susceptible to illumination variations and noises. We tackle this problem by fusing the data of these two sensors to complement each other's disadvantages. To achieve better fusion, the LiDAR data is transformed to image-like data and LiDAR features are also transformed adaptively. For better segmentation, we keep high resolution features throughout the convolutional network to reduce information loss and improve segmentation precision. The LiDAR-camera fusion are incorporated into the high-resolution network at multiple layers and multiple scales to constitute our road segmentation system. Comprehensive experiments on KITTI road dataset have been conducted to verify the effectiveness and efficiency of the proposed method.