Intelligent detection and filtering of swarm noise from drone acquired LiDAR data using PointPillars

We present an approach to removing swarm selfnoise from airborne LiDAR data using the point-based PointPillars deep learning neural network (DNN) which was trained to detect and localize drones. The use of hyperlocalized swarms of survey drones can improve the productivity of maintenance inspection operations, with trajectory-based mission planning capable of mitigating air-to-air collision risk. However, even though they are strategically separated, individual drones often collect survey data that is cluttered with frequent observations of the rest of the swarm. This paper describes the proposed denoising method which was tested using LiDAR survey data collected during an inspection of a coastal railway bridge. The DNN performs favorably with respect to classical radius and statistical filtering methods. We show that a combined approach of the DNN algorithm and classical methods provides the best results, successfully removing over 99% of swarm self-noise and without any false positives when applied to a 7-million-point LiDAR dataset.