Enhancing Aerial Camera-LiDAR Registration through Combined LiDAR Feature Layers and Graph Neural Networks

Abstract. Integrating optical images with Light Detection and Ranging (LiDAR) data is an important advance in Photogrammetry, Geomatics and Computer Vision, registering the strengths of both modalities (height and spectral information). Most orthoimages and aerial LiDAR data are georeferenced to a common ground coordinate system; however, a registration gap remains, and achieving high-accuracy registration between these datasets is challenging due to their differing data formats and frames of reference. In this paper, we propose an approach to enhance camera-LiDAR registration through combined LiDAR feature layer generation and Deep Learning. Our method involves creating weighted combinations of feature layers from LiDAR data, leveraging intensity, elevation, and bearing angle attributes. Subsequently, a 2D-2D Graph Neural Network (GNN) pipeline serves as an intermediate step for feature detection and matching, followed by a 2D-3D affine transformation model to register optical images to point clouds. Experimental validation across aerial scenes demonstrates significant improvements in registration accuracy. Notably, in urban building areas, we achieved an RMSE of around 1.1 pixel, marking a reduction of 5 pixels compared to georeferenced baseline values. In rural road scenes, our method yielded a pixel RMSE of 1.3, with a 4-pixel reduction compared to baseline results. Additionally, in water scenes, which tend to be noisy in LiDAR data, we achieved a pixel RMSE of 1.8, representing a slight half-pixel reduction compared to the baseline. Therefore, by using weighted and combined LiDAR feature layer and GNN feature matching, this approach augments the number of key points and matches, directly correlating with the observed registration reduction in pixel RMSE across diverse aerial scene types.