Raw fusion of camera and sparse LiDAR for detecting distant objects

Abstract Environment perception plays a significant role in autonomous driving since all traffic participants in the vehicle’s surroundings must be reliably recognized and localized in order to take any subsequent action. The main goal of this paper is to present a neural network approach for fusing camera images and LiDAR point clouds in order to detect traffic participants in the vehicle’s surroundings more reliably. Our approach primarily addresses the problem of sparse LiDAR data (point clouds of distant objects), where due to sparsity the point cloud based detection might become ambiguous. In the proposed model each 3D point in the LiDAR point cloud is augmented by semantically strong image features allowing us to inject additional information for the network to learn from. Experimental results show that our method increases the number of correctly detected 3D bounding boxes in sparse point clouds by at least 13–21 % and thus raw sensor fusion is validated as a viable approach for enhancing autonomous driving safety in difficult sensory conditions.