基于三维激光雷达的实时目标检测

发布时间：2018-01-13 14:24

本文关键词：基于三维激光雷达的实时目标检测　出处：《浙江大学》2014年硕士论文　论文类型：学位论文

【摘要】：无人自主车是一种集环境感知、规划决策、运动控制功能于一体的人工智能平台。无人车对周围环境的感知需求是多方面的,其中之一就是对障碍物进行检测,并在此基础上,对感兴趣的目标,如车辆和行人,进一步进行检测和判别,以便合理应对。在可使用的传感器中,激光雷达由于其高精度测距、探测范围大、抗干扰能力强等优点,近年来在移动机器人上的应用越来越广泛。因此本文研究的重点是基于三维激光雷达的目标检测。本文将目标检测分为障碍物检测和目标分类两个阶段。对于障碍物检测,首先介绍了论文研究的64线激光雷达传感器,并推导了三维点云坐标转化的内参模型和外参模型。采用基于障碍栅格的检测方法,将三维点云进行栅格投影和属性判断。为了抑制传感器和环境噪声,进行悬空点滤除和单点滤除。障碍栅格检测具有快速稳定,缺点是对数据的分割精度较低,易产生欠分割。因此本文提出了基于点云梯度的局部最优分割方法,首先对扫描线的点进行梯度分割,然后在聚类的分割段中依据障碍属性滤除误检,最后根据可靠的非障碍点估计局部地面来恢复漏检点。实验表明基于点云梯度的局部最优分割方法效果比栅格方法好,但由于其计算量较大,实时性受到限制。在得到障碍物信息的基础上,采用基于距离相关性的聚类和提取障碍物轮廓,并讨论了障碍的可通行区域检测。对于目标分类,本文将环境中的障碍物分为三类：车辆、行人、其它。检测过程中自主车是运动的,因此无论障碍物是否运动,在自主车看来都是运动的,所以本文不从是否运动的角度来分类障碍物。这样做的好处是即使车辆或者行人是静止的,也能得到较好的检测。分类的主要思想是先检测车辆和非车辆,然后在非车的障碍物中根据简单几何特征来判断行人。针对车辆几何形状相似难以区分和物体遮挡等问题,提出了三种新特征描述：反射强度概率分布、纵向高度轮廓分布和位置姿态相关特征。最后根据障碍物的三维点云数据特征,实现了基于SVM分类器的实时车辆检测。在城市复杂路口下实验,验证了本文所提车辆和行人检测特征的有效性,并且整体系统能有较好的目标检测精度和实时性能。
[Abstract]:Unmanned vehicle (UAV) is an artificial intelligence platform with environmental perception, planning, decision making and motion control functions. There are many perceptual requirements for the surrounding environment of UAV, one of which is to detect obstacles. And on this basis, the interested targets, such as vehicles and pedestrians, further detection and discrimination, in order to reasonably deal with. In the usable sensors, lidar has a wide range of detection due to its high accuracy ranging. In recent years, the anti-jamming ability is more and more widely used in mobile robot, so the focus of this paper is the target detection based on 3D lidar. In this paper, target detection is divided into two stages: obstacle detection and target classification. For obstacle detection, the 64-line lidar sensor studied in this paper is first introduced. The internal parameter model and external parameter model of 3D point cloud coordinate transformation are derived. Using the detection method based on barrier grid, the 3D point cloud is projected and attribute judgment is carried out in order to suppress sensor and environmental noise. The obstacle grid detection is fast and stable, the disadvantage is that the data segmentation accuracy is low, and it is easy to generate under-segmentation. Therefore, a local optimal segmentation method based on point ladder degree is proposed in this paper. Firstly, the points of the scan line are divided by gradient, and then the false detection is filtered according to the obstacle attributes in the segmentation segment of the clustering. Finally, the local surface is estimated according to the reliable non-obstacle points to recover the missing points. The experimental results show that the local optimal segmentation method based on the point ladder degree is better than the grid method, but because of the large amount of calculation. On the basis of obtaining obstacle information, the obstacle contour is extracted by clustering based on distance correlation, and the tradable area detection of obstacle is discussed. For target classification, the obstacles in the environment are divided into three categories: vehicles, pedestrians, others. In the process of detection, autonomous vehicles are moving, so no matter whether the obstacles are moving or not, they are all moving in the view of autonomous vehicles. Therefore, this paper does not classify obstacles from the point of view of motion. The advantage of this method is that even if the vehicle or pedestrian is stationary, it can also get better detection. The main idea of classification is to detect vehicles and non-vehicles first. Then the pedestrian is judged according to the simple geometric features in the non-vehicular obstacles. In view of the difficulty of distinguishing the geometric shape similarity of the vehicle and the object occlusion, three new features are proposed: the probability distribution of reflection intensity. Finally, according to the 3D point cloud data features of obstacles, real-time vehicle detection based on SVM classifier is realized. The effectiveness of the proposed vehicle and pedestrian detection features is verified, and the overall system has good target detection accuracy and real-time performance.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN958.98

【共引文献】