GPU-Oriented Environmental Cognition of Power Transmission Lines Through LiDAR-Equipped UAVs

This article proposes an environmental cognition method for execution on a graphics processing unit (GPU) oriented single-board computer (SBC). We develop a lightweight environmental cognition system for real-time mapping by an unmanned aerial vehicle (UAV) by considering the weight of a cognition system as a tradeoff between the flight time and mapping efficiency and loss of environmental cognition capabilities. Heavy systems consume battery power rapidly while ensuring high computational performance and impacting flight envelope. The proposed method enhances real-time mapping speed by using GPU parallelism and minimizing the data transfer between the embedded central processing unit (CPU) and GPU. Specifically, the point cloud data (PCD) from the light detection and ranging are transformed into global coordinates and voxelized. The occupancies of the voxels are updated in a probabilistic manner to eliminate dynamic noise. The analysis of field tests confirms that the proposed method generated and updated the voxel map in real time without losses. In contrast, Octomap executed on a CPU- or GPU-oriented SBC generated and updated the voxel map in a limited manner, resulting in the significant loss of PCD due to the computational burden or heavy data transfer traffic from GPU to CPU. The proposed method contributes to develop a smart environmental cognition system for the autonomous flight of UAVs.