Parallel Processing of 3D Object Recognition by Fusion of 2D Images and LiDAR for Autonomous Driving

At the moment, autonomous driving requires a lot of sensors: cameras, lidar, etc. It takes a lot of time and resources to process all the input data from these sensors. In this paper, we reduce the processing time and resources of lidar and camera data by parallelizing the input data of autonomous vehicles. Cameras mounted on autonomous vehicles are often wide-angle or have multiple angles of view. These multiple camera inputs are flattened and processed in parallel, and then YOLO is used to combine the 3D data from the lidar with the 2D inputs from the camera. By combining cameras from multiple angles and processing them in parallel, except where they overlap, you can reduce the time it would take to process each image serially. This algorithm is also highly scalable as it can be applied to a single camera rather than multiple camera sensors. Experiments were conducted using KITTY and YOLO with labelled 3D lidar data and 2D image data. The FPS is 7.98, which is fast, and the parallel processing reduces the time by about 1.4 times.