无人驾驶汽车动态障碍物避撞关键技术研究

发布时间：2018-08-17 19:25

【摘要】：随着汽车保有量的增加,交通拥堵和交通事故发生率不断升高,作为解决此问题的重要手段,无人驾驶汽车的研究日益迫切。无人驾驶汽车在复杂交通环境中行驶时会不可避免地与其他交通参与者如汽车、行人和自行车等产生交互,在交互过程中必须避开所有潜在碰撞以保证行驶安全。为了完成此任务,首先需要无人驾驶汽车能够准确地检测并跟踪动态障碍物,估计其运动状态;其次,不同种类的动态障碍物具有不同的运动特性,为提高避撞行为的合理性,需要识别出动态障碍物的种类以便无人驾驶汽车执行更加合理的避撞行为;最后为了避开与动态障碍物之间的潜在碰撞,要求无人驾驶汽车能够准确预测出动态障碍物的运动轨迹,尤其是运动速度较快的动态车辆。但现有的基于单一轮廓特征的方法在动态障碍物检测跟踪过程中的准确率和速度较低,无法满足动态障碍物避撞安全性的要求,基于轮廓特征和运动状态的动态障碍物识别算法的准确率较低且识别范围较小,无法满足无人驾驶汽车避撞合理性的要求,动态车辆行驶轨迹由很多因素决定,现有的基于动态车辆实时运动状态的轨迹预测方法误差较大,不满足无人驾驶汽车避撞准确性的要求。针对以上问题,本文提出了基于多特征融合的动态障碍物检测跟踪方法、基于时空特征向量的动态障碍物识别方法和基于驾驶行为意图检测的动态车辆轨迹预测方法,从而实现更加安全、合理和准确的动态障碍物避撞,主要研究内容如下:1)动态障碍物的检测跟踪:为了提高动态障碍物检测跟踪的准确性与速度,提出了一种基于多特征融合的动态障碍物检测跟踪方法。首先分别从三维激光雷达获得的数据和多层激光雷达获得的数据中提取障碍物的轮廓特征和激光脉冲反射强度特征,接着融合提取的特征并对动态障碍物进行建模,通过构建相似度矩阵完成动态障碍物的匹配跟踪并利用建立的障碍物模型完成动态障碍物的运动状态估计,为动态障碍物识别和动态车辆轨迹预测提供障碍物运动状态信息。2)动态障碍物的识别:不同类型的动态障碍物具有不同的运动特性且需要不同的避撞策略,为了使无人驾驶汽车能够更加合理地选择避撞行为,本文提出了一种基于时空特征向量的无人驾驶汽车动态障碍物识别方法。首先结合障碍物空间维度上的几何轮廓特征、时间维度上的Zernike不变矩特征和无人驾驶汽车的位姿信息构建时空特征向量,在此基础上使用AdaBoost算法构建动态障碍物分类器,最后使用分类器识别动态障碍物,为避撞行为决策提供障碍物类别信息。3)动态车辆轨迹预测:在动态障碍物识别的基础上,针对其中动态车辆轨迹预测不准确的问题,提出了基于驾驶行为意图检测的轨迹预测方法。首先,利用高斯混合模型从车辆驾驶行为数据和道路结构信息中学习驾驶行为模式并用于检测动态车辆的驾驶行为意图,接着根据驾驶行为意图检测结果计算动态车辆的长期理想轨迹,并结合运动模型预测动态车辆的行驶轨迹,为准确的碰撞检测和避撞奠定基础。最后,通过在城区交通环境中的实车实验验证了上述技术的可靠性与稳定性,极大地提升了动态障碍物检测跟踪的准确性与速度,提高了动态障碍物的识别准确率和动态车辆的轨迹预测精度并扩展了动态障碍物识别距离范围。
[Abstract]:With the increase of vehicle ownership, the incidence of traffic congestion and traffic accidents is increasing. As an important means to solve this problem, the study of driverless vehicles is becoming more and more urgent. In order to accomplish this task, it is necessary to detect and track dynamic obstacles accurately and estimate their motion state. Secondly, different kinds of dynamic obstacles have different motion characteristics. In order to improve the rationality of collision avoidance behavior, it is necessary to identify them. Finally, in order to avoid the potential collision with the dynamic obstacles, the unmanned vehicle is required to accurately predict the trajectory of the dynamic obstacles, especially the fast moving vehicles. The method has low accuracy and speed in the process of dynamic obstacle detection and tracking, and can not meet the requirements of dynamic obstacle avoidance safety. The dynamic obstacle recognition algorithm based on contour features and motion state has low accuracy and small recognition range, and can not meet the requirements of the rationality of unmanned vehicle collision avoidance. Vehicle trajectory is determined by many factors. Existing trajectory prediction methods based on dynamic vehicle real-time motion state have large errors and do not meet the requirements of collision avoidance accuracy of unmanned vehicles. Recognition method and dynamic vehicle trajectory prediction method based on driving behavior intention detection are used to achieve more safe, reasonable and accurate dynamic obstacle avoidance. The main research contents are as follows: 1) Dynamic obstacle detection and tracking: In order to improve the accuracy and speed of dynamic obstacle detection and tracking, a multi-feature fusion method is proposed. A combined dynamic obstacle detection and tracking method is proposed. First, the outline features and laser pulse reflective intensity features of obstacles are extracted from the data obtained by three-dimensional lidar and multi-layer lidar respectively, then the extracted features are fused and the dynamic obstacles are modeled, and the dynamic obstacles are completed by constructing similarity matrix. Dynamic obstacle recognition: Different types of dynamic obstacles have different motion characteristics and require different collision avoidance strategies for dynamic obstacle recognition and dynamic vehicle trajectory prediction. In order to make the unmanned vehicle choose the collision avoidance behavior more reasonably, this paper proposes a dynamic obstacle recognition method for unmanned vehicle based on space-time feature vector. Firstly, the geometric contour feature of obstacle space dimension, the Zernike invariant moment feature of time dimension and the pose information construction of unmanned vehicle are combined. On the basis of space-time feature vectors, AdaBoost algorithm is used to construct a dynamic obstacle classifier. Finally, the classifier is used to identify the dynamic obstacles and provide obstacle class information for collision avoidance decision-making. 3) Dynamic vehicle trajectory prediction: On the basis of dynamic obstacle recognition, aiming at the problem of inaccurate dynamic vehicle trajectory prediction, the dynamic vehicle trajectory prediction is proposed. A trajectory prediction method based on driving behavior intention detection is proposed. Firstly, driving behavior patterns are learned from driving behavior data and road structure information by using Gaussian mixture model and used to detect driving behavior intention of dynamic vehicles. Finally, the reliability and stability of the above techniques are verified by the real vehicle experiments in urban traffic environment, which greatly improves the accuracy and speed of dynamic obstacle detection and tracking, and improves the accuracy of dynamic obstacle identification. Accuracy and dynamic vehicle trajectory prediction accuracy and extend the range of dynamic obstacle recognition.
【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：U463.6

【相似文献】