3D Lidar Based Real-time Obiects Detection

Unmanned autonomous vehicle is composed of environmental awareness, planning decisions, motion control functions, which is regarded as one of the artificial intelligence platform. The perception needs of surrounding environment in unmanned autonomous vehicle is multifaceted, one of which is for obstacle detection, and on that basis, the target of interest such as vehicles and pedestrians, further detection and identification to respond reasonable. Among so-far available sensors, due to laser radar of high-precision distance, strong anti-jamming ability, etc. Recent years, applications in mobile robot are more and more widely. So in this thesis we focus on study about target detection using three-dimensional Lidar.This thesis we divide target detection into two phases:obstacle detection and target classification. For obstacle detection, we first introduce the sensor Velodyne HDL-64E, and a detailed derivation of three-dimensional coordinate transformation mode:the intrinsic and extrinsic calibrate model. Then we use occupancy grid map method to detect obstacle, the grid will be projected by three-dimensional point clouds and determine the grid’s properties. In order to suppress environmental and sensor noise, we use floating and single points filter. Occupancy grid map method has fast and stable results, but with lower accuracy and prone to cause under-segmentation. Therefore, we proposed Ring Gradient based Local Optimal Segmentation (RGLOS): first we computer the gradient scanning point clouds to segment all points; second we filter mistaken detected points in cluster segmentation according to obstacle property; last using reliable non-obstacle points to estimate local ground height to recover omit obstacle points. The experimental results indicate that the method of RGLOS gradient-based segmentation method is better than occupancy grid map method. Due to its large amount of calculation, the method has restrictions in real-time. Last, based on the information of obtained obstacle, we cluster obstacle by correlation of distance and extract object contour, and discusses accessible area detection. For target classification, we divide the environment objects into three types: vehicles, pedestrians, others. Autonomous vehicle is usual in motion, so no matter whether the obstacles move or not in autonomous vehicles will seem movement. Considering that, we do not rely on movement features to classify object. The advantage of that is even if the target object is stationary or not, we can also get a better detection. The main idea is first to classify vehicles and non-vehicles, and in the non-car object we use method based on simple geometric features to determine whether pedestrians or not. To solve difficulty to distinguish objects’geometry similar and occlusion issues, we proposed three novel features:intensity reflection probability distribution, lengthways contour in height, and position-related features. A support vector machine (SVM) classifier was used to realize real-time detecting vehicle. Under real complex urban environment experiment results demonstrate the effectiveness of vehicle and pedestrian detection features mentioned in this thesis. And the whole system can have a better accuracy and real-time target classification.