3D Vehicle Detection Using Cheap LiDAR and Camera Sensors

Autonomous Vehicles (AVs) are expected to be intelligent enough to perceive the world accurately in terms of avoiding road obstacles. Remarkable progress has been made in 3D road scene perception of AVs through machine learning and computer vision methods, but existing solutions rely on expensive 64 beams LiDAR point clouds for the 3D positioning of objects. In this paper, we propose a simple yet effective approach that is based on the success of 2D object detection to estimate 3D positions of the vehicles in front of AVs. Our approach relies on camera RGB images for predicting size and orientation of 3D bounding boxes of AVs by using a novel deep neural network (DNN) and LiDAR 3D point clouds for distance estimation. For testing and training, KITTI and Waymo datasets are employed. We have converted 64 beams of LiDAR point clouds into 32 and 16 beams point clouds for model performance analysis. Based on the results, the proposed method proved to be robust with sparse point clouds without compromising accuracy.