LiDAR Based Real Time Fast Object Detection and Tracking on UAV Using VoxelNet

Real-time object detection and tracking on Unmanned Aerial Vehicles (UAVs) is a challenging task due to dynamic flight conditions, limited onboard resources, and the need for accurate perception in complex environments. Traditional vision-based systems often struggle in low-light or cluttered scenes, limiting their reliability. To address challenges, this paper presents a real-time LiDAR-based object detection and tracking system for UAVs using the Voxel Net deep learning architecture. The proposed system begins by filtering raw LiDAR point cloud data to remove outliers, such as points too distant from the sensor. Ground points are effectively removed using a RANSAC-based algorithm, followed by segmentation into ground and nonground components. Both segments are then clustered to identify distinct objects in the scene. Voxel-Net is employed to detect and track these objects in 3D by generating bounding boxes with spatial and depth information. The proposed system achieves high efficiency with processing and filtering in 8 ms, segmentation in 8 ms, and clustering in 12 ms, making it suitable for fast and accurate object tracking in UAV-based real-time applications.