Yu, Y., Gao, Z., Zhu, B., and Zhao, J., "Recognition and Classification of Vehicle Target Using the Vehicle-Mounted Velodyne LIDAR," SAE Technical Paper 2014-01-0322, 2014, doi:10.4271/2014-01-0322.
This paper describes a novel recognition and classification method of vehicle targets in urban road based on a vehicle-mounted Velodyne HDL64E light detection and ranging (LIDAR) system. The autonomous vehicle will choose different driving strategy according to the surrounding traffic environments to guarantee that the driving is safe, stable and efficient. It is helpful for controller to provide the efficient stagey to know the exact type of vehicle around. So this method concentrates on reorganization and classification the type of vehicle targets so that the controller can provide a safe and efficient driving strategy for autonomous ground vehicles.The approach is targeted at high-speed ground vehicle, so real-time performance of the method plays a critical role. In order to improve the real-time performance, some methods of data preprocessing should be taken to simplify the large-size long-range 3D point clouds. First, given the large amount of data delivered by 360° range scanners, the most efficient method to date is reducing scale, so the valid district can be extracted and the point cloud is compressed. Second, the targets among the amount of points should be extracted. The algorithms is to separate the targets from the ground points. Consequently, the ground points can be removed and then the remaining points are distributed into separate clusters according to the distance of each point. Each cluster count as a vehicle-like target. This section of data processing is proved to be effective and efficient by experiment.The vehicle target should be recognized from the vehicle-like target using border extraction method from range imagines. Range-imagine can be achieved by projecting the 3D points to a 2.5D grid and taking the LIDAR (Light Detection and Ranging) origin point as the project origin. In this method, the transform just uses in the each cluster instead of whole 3D points. By reducing dimensionality, the border can figure out efficient. In different with the common method, this paper not only presents how to recognize vehicle from vehicle-like target, but how to classify the type of the vehicle. So several possible categories have been identified based on the border features. And this detailed division requires sufficient training data for each specific class. Through the adequate experiment, an ideal result is achieved.