复杂环境下智能车辆动态目标三维感知方法研究

发布时间：2019-01-21 19:31

【摘要】：驾驶员对交通环境信息处理包括感觉、感知和认知三个阶段,同样,智能车辆“驾驶脑”只有像人类一样达到对环境感知的升华,才能做到真正的“智能”驾驶。作为智能车提高环境感知深度的重要手段,当前基于三维激光雷达的动态目标感知研究已获得极大的关注并取得许多成果。但仍存在以下问题:一是感知的环境简单,即通常集中于单一感知对象、稀疏目标分布的简单交通场景,较少对目标混杂、密集分布的复杂场景进行研究;二是多处于低层次感知研究阶段,即以动态目标的检测及识别为研究重点,较少深入认知目标内在行为。针对上述问题,本文对复杂环境下基于三维激光雷达的动态目标感知方法进行研究,围绕动态目标的感觉、感知及认知机理,以其状态、类别及行为感知为具体研究对象,并通过自然环境下采集的数据对系统性能进行验证。论文首先立足于智能车辆环境感知的两个科学问题,对动态目标感知的研究对象和范围进行明确,提出从动态目标状态、类别至行为的从低级至高级感知的研究思路,并在此基础上构建总体研究方案及体系结构。针对于海量点云条件下目标检测准确性与实时性的矛盾,将点云分割转换至图像域中并利用图像学中成熟方法来进行处理。而后,为克服目标识别中不同距离下目标点云密度差异性问题,提出一种点云密度增强方法来生成均匀、满足密度要求的新的目标点云,以提高点云特征的适用性及模型分类能力。在此基础上,对车辆目标关键状态参数进行建模,采用决策树-跟踪门及联合概率数据关联-卡尔曼滤波的跟踪方法来进行参数优化。最后对车辆目标运动行为进行分析建模,提出了基于自然语音识别方法的车辆行为识别模型,并以换道行为作为对象进行实现。为验证方法有效性,以自然交通环境下采集数据进行实验验证。结果证明,本文研究三维感知系统能够在复杂交通环境下快速、准确地计算出动态目标的状态及类别参数,同时基于自然语音识别方法的行为识别模型不仅有助于建立起通用的车辆行为识别框架和体系,其输出结果还能准确、有层次的体现行为程度,提高感知深度。
[Abstract]:Drivers' processing of traffic environment information includes three stages: sensation, perception and cognition. Similarly, intelligent vehicle "driving brain" can only achieve real "intelligent" driving if it is sublimated to environmental perception like human beings. As an important means for intelligent vehicle to improve the depth of environmental perception, the research of dynamic target perception based on 3D lidar has been paid great attention to and many achievements have been made. However, the following problems still exist: first, the environment of perception is simple, that is, the simple traffic scene is usually focused on a single perceptual object, sparse target distribution, and less research on the complex scene where the target is mixed and densely distributed; The second is that most of them are in the stage of low-level perception, that is to say, the focus of the research is on the detection and recognition of dynamic targets, and less on the intrinsic behavior of cognitive targets. In view of the above problems, this paper studies the dynamic target perception method based on 3D lidar in complex environment, focusing on the perception, perception and cognitive mechanism of dynamic target, taking its state, category and behavior perception as the specific research object. The performance of the system is verified by the data collected in the natural environment. Firstly, based on the two scientific problems of intelligent vehicle environment perception, the research object and scope of dynamic target perception are clarified, and the research ideas from dynamic target state, category to behavior from low-level to advanced perception are put forward. On this basis, the overall research scheme and architecture are constructed. Aiming at the contradiction between accuracy and real-time of target detection under the condition of massive point cloud, the segmentation of point cloud is transformed into image domain and processed by mature methods in image science. Then, in order to overcome the difference of point cloud density at different distances in target recognition, a point cloud density enhancement method is proposed to generate a new target point cloud that meets the requirements of target density. In order to improve the applicability of point cloud features and model classification ability. On this basis, the key state parameters of vehicle target are modeled, and the parameters are optimized by using the tracking method of decision tree-tracking gate and joint probabilistic data association Kalman filter. Finally, a vehicle behavior recognition model based on natural speech recognition is proposed, which is based on the analysis and modeling of vehicle moving behavior. In order to verify the effectiveness of the method, the data collected under the natural traffic environment were tested. The results show that the 3D sensing system can quickly and accurately calculate the state and class parameters of dynamic targets in complex traffic environment. At the same time, the behavior recognition model based on natural speech recognition method can not only help to establish a general framework and system of vehicle behavior recognition, but also can accurately and hierarchically reflect the degree of behavior and improve the depth of perception.
【学位授予单位】：清华大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：U463.6;TN958.98

【相似文献】