考虑交通车辆运动不确定性的轨迹规划方法研究

发布时间：2019-06-05 17:15

【摘要】：汽车智能化是应对汽车工业发展所面临的安全、拥堵和环保等诸多问题的关键技术途径,也是汽车技术发展的必然趋势。做为智能车辆的关键技术之一,轨迹规划需要对规划状态进行准确地危险评估,并基于此规划出车辆的行驶路径和速度,从而保证智能车辆在交通环境中的行驶安全性。这要求在轨迹规划中必须对交通车辆的运动轨迹做出合理的预测。但对智能车辆而言,交通车辆的未来运动是不确定的,具有一定的随机性。忽略交通车辆运动的不确定性将导致危险评估结果不够准确,从而影响智能车辆的行驶安全性。因此,在对交通车辆进行轨迹预测时,不仅不可忽略运动的不确定性,还必须获取其准确的概率特性。与此同时,确定性的危险评估结果也已经不能准确反映规划状态的安全性,其安全性仅能以碰撞概率的形式表达。为了提高轨迹规划的性能、保障智能车辆的行驶安全性,必须充分考虑由交通车辆运动不确定性引起的碰撞概率的影响。基于行为的运动模型框架是预测交通车辆运动轨迹的有效方法。但驾驶人不同的驾驶风格使得同一驾驶行为下的运动轨迹有着不同的运动模式。若忽略该差异必然导致预测所得概率特性不够准确。因此,为了提高预测准确性需建立不同模式的运动模型并实现运动模式的辨识。基于支持向量机的分类器是解决辨识问题的有效方法。传统的分类器将输入样本视为独立存在的个体,其结果依赖于分类器自身性能以及当前输入样本。但由于难以通过车载传感获取交通车辆内部参数以及驾驶员状态、车辆状态的实时数据,使得对模式辨识仅能依赖有限的外部传感信息。因此,难以保证单分类器对单样本辨识结果的准确性。在基于行为的运动模型框架下,高斯过程运动模型是描述汽车运动随机性的有效方法,建立不同运动模式所对应的运动模型是实现交通车辆轨迹预测的基础。但直接以运动模型表征交通车辆运动不确定性的概率特性并不准确,必须考虑模型中与实时运动轨迹相匹配的先验向量对预测向量概率特性的影响。而运动模式辨识仅确定了实时轨迹的运动模型,与之匹配的先验向量依然是未知的,现有研究中对该问题的解决鲜有提及。快速搜索随机树具有概率完备性及快速求得可行解的能力,是解决汽车轨迹规划问题的有效方法。传统研究多集中于对汽车运动学、动力学约束及算法实时性等问题的考虑,较少关注交通车辆运动不确定性的影响。因此,传统方法通常以逻辑判断的形式表达搜索树中节点的安全性,并基于此实现轨迹的搜索与决策。而交通车辆运动的不确定性使得上述条件不再成立。所以,即便通过碰撞概率实现对规划状态准确地危险评估,传统方法的规划机理依然不能有效地处理碰撞的概率性对搜索过程的影响,使得对不确定性的处理中一定的盲目性,从而难以保证规划方法的性能及智能车辆的行驶安全性。针对当前研究中的不足,本文对考虑交通车辆运动不确定性的轨迹规划问题进行了研究,主要研究内容如下:第一,本文提出了一种交通车辆运动模式辨识方法。方法建立了基于“一对多”纠错输出码的辨识架构,将运动模式辨识问题转化为若干二分类问题。随后通过成对比较分析,建立了概率估计模型并以最小相对熵为优化目标完成实际概率的估计,从而以多分类器取代单分类器实现对样本的辨识。建立了贝叶斯推理模型以揭示连续若干概率估计结果与最终辨识模式结果的关系,从而以多样本取代单样本实现对运动模式的辨识。本文所述方法与传统方法的实验对比结果表明,所述方法能够有效的消除单分类器以及单样本带来的错误辨识结果,从而有效地提高辨识的准确性。第二,本文提出了基于高斯过程运动模型的轨迹预测方法。首先完成运动轨迹的模式聚类并建立基于高斯过程的运动模型。在进行轨迹预测时,提出并推导了基于马氏距离的先验向量计算方法,从而有效的建立实时运动轨迹与运动模型之间的匹配关系。随后推导了基于条件高斯分布的轨迹预测方法以获得交通车辆未来运动轨迹的概率特性。实验结果表明,所述方法能够准确计算出先验向量的维数,从而保证预测所得概率特性的准确性。第三,本文提出一种考虑不确定性的轨迹规划方法。方法包括考虑汽车行驶环境特性的采样策略以及考虑汽车运动特性的节点距离度量策略。为了在轨迹规划过程中考虑由交通车辆运动不确定性引起的碰撞的概率性的影响,方法以碰撞概率表达对规划状态准确的危险评估,并将其建模为搜索树中节点的代价,从而在规划机理中考虑碰撞概率性的影响。以此为基础开展搜索树节点排序、采样节点扩张、目标偏向扩张以及轨迹评价与决策,能够保证不确定性下规划方法的性能。最后,开展了多种工况下的仿真实验:单步对比实验表明,考虑不确定性的轨迹规划可以实现更为准确的危险评估从而决策出更为安全的轨迹;动态避障实验表明,基于准确的危险评估,规划过程中搜索树朝向更为安全的区域扩张,有效地消除了不确定性下规划的盲目性、随机性。本文基于国家自然科学基金重点项目搭建了实车实验平台以验证本文研究内容和提出的方法。首先,开展了包括基于单点预瞄的路径跟随控制器设计以及基于预瞄加速度的速度跟随控制器设计的研究工作。其次,为了实现实车实验验证,完成了以某型轿车为实车平台的总体方案设计,包括平台软硬件架构设计,机械、通讯及供电系统设计,最终改装并完成了实车平台搭建。接着,根据实车实验需求,对轨迹规划、轨迹跟随控制算法中所涉及的平台参数进行了估计。最后,依托所搭建的平台中对本文所涉及方法进行了验证。实验结果表明,本文所述方法能够有效地提高智能车辆轨迹规划的准确度,进而更好地保证其行驶安全性,验证了本文所述方法的有效性。
[Abstract]:The intelligent of automobile is the key technology to deal with the problems such as safety, congestion and environmental protection in the development of automobile industry. It is also an inevitable trend of the development of automobile technology. As one of the key technologies of the intelligent vehicle, the track planning needs to carry out the accurate risk assessment of the planning state, and the running route and the speed of the vehicle are planned based on the planning state, thereby ensuring the running safety of the intelligent vehicle in the traffic environment. This requires that a reasonable prediction of the motion trajectory of the traffic vehicle must be made in the trajectory planning. However, in the case of intelligent vehicles, the future movement of the traffic vehicle is uncertain and has a certain randomness. Ignoring the uncertainty of traffic vehicle motion will result in a lack of accuracy in the risk assessment, which will affect the travel safety of the smart vehicle. Therefore, it is not only important to ignore the uncertainty of the movement, but also to obtain the accurate probability characteristic of the traffic vehicle. At the same time, the outcome of the risk assessment of certainty has also been unable to accurately reflect the safety of the planned state, and its safety can only be expressed in the form of collision probability. In order to improve the performance of the trajectory planning and to guarantee the running safety of the intelligent vehicle, the impact of the collision probability caused by the motion uncertainty of the traffic vehicle must be fully considered. The behavior-based motion model framework is an effective method for predicting the motion track of a traffic vehicle. But the driver's different driving styles have different modes of motion under the same driving behavior. If that difference is ignore, the probability characteristic of the prediction is not accurate enough. Therefore, in order to improve the accuracy of the prediction, it is necessary to establish a motion model of different modes and to realize the identification of the motion pattern. A classifier based on support vector machine is an effective method to solve the problem of identification. Traditional classifiers treat the input samples as individuals that are independent, and the results depend on the classifier's own performance and the current input samples. However, since it is difficult to acquire that internal parameters of the traffic vehicle and the driver state and the real-time data of the vehicle state by the on-vehicle sensor, the mode identification can only depend on the limited external sensing information. Therefore, it is difficult to ensure the accuracy of the single-classifier to the single-sample identification result. In the framework of the behavior-based motion model, the Gaussian process motion model is an effective method to describe the randomness of the motion of the automobile, and the motion model corresponding to the establishment of the different motion modes is the basis for realizing the track prediction of the traffic vehicle. However, the probability characteristic of the motion uncertainty of the traffic vehicle is not accurately characterized by the motion model, and the influence of the prior vector matching the real-time motion track on the probability of the prediction vector must be taken into consideration. And the motion pattern recognition only determines the motion model of the real-time track, the prior vector matched with the motion model is still unknown, and the solution to the problem in the prior research is rarely mentioned. The ability of fast search of random tree with probability completeness and fast finding feasible solution is an effective method to solve the problem of automobile track planning. The traditional research focuses on the problems such as kinematics, dynamics and real-time performance of the vehicle, and is less concerned with the influence of the traffic uncertainty of the traffic vehicles. Therefore, the traditional method generally expresses the security of the nodes in the search tree in the form of a logical judgment, and realizes the search and decision-making of the trajectory based on the method. And the uncertainty of the movement of the traffic vehicle makes the above-mentioned conditions no longer established. Therefore, even though the risk assessment of the planning state is accurately carried out by the collision probability, the planning mechanism of the traditional method can not effectively deal with the impact of the probability of the collision on the search process, so that the blindness in the processing of the uncertainty is certain, Thus it is difficult to guarantee the performance of the planning method and the running safety of the intelligent vehicle. In view of the deficiency in the current research, this paper studies the trajectory planning of the vehicle motion uncertainty, and the main contents are as follows: First, this paper presents a method for identifying the motion pattern of the traffic vehicle. Methods The recognition architecture based on the "a pair of" error correction output code was established, and the problem of identification of motion pattern was transformed into a number of two-class problems. Then, based on the comparative analysis, the probability estimation model is established and the estimation of the actual probability is completed with the minimum relative entropy as the optimization target, so that the identification of the samples is realized by the multi-classifier instead of the single classifier. A Bayesian inference model is established to reveal the relationship between the successive probability estimation results and the result of the final identification mode, so as to realize the identification of the motion mode with a variety of the substitute single samples. The experimental results of the method and the traditional method show that the method can effectively eliminate the error identification result caused by the single classifier and the single sample, thereby effectively improving the accuracy of the identification. Secondly, the paper presents a method of trajectory prediction based on the Gaussian process motion model. Firstly, the mode clustering of the motion track is completed and a motion model based on the Gaussian process is established. In the course of trajectory prediction, a priori vector calculation method based on the Markovian distance is proposed and the matching relation between the real-time motion trajectory and the motion model is effectively established. The trajectory prediction method based on conditional Gaussian distribution is then derived to obtain the probability characteristic of the future motion trail of the traffic vehicle. The experimental results show that the method can accurately calculate the dimension of the prior vector, so as to ensure the accuracy of the probability characteristic of the prediction. Thirdly, this paper puts forward a method of trajectory planning which takes into account the uncertainty. The method comprises the following steps: taking into account the sampling strategy of the vehicle running environment characteristic and the node distance measurement strategy taking into account the automobile motion characteristic. in order to consider the influence of the probability of the collision caused by the motion uncertainty of the traffic vehicle in the course of the trajectory planning, the method expresses the risk assessment of the planning state with the collision probability expression, and the model is the cost of the node in the search tree, So that the influence of the collision probability is taken into consideration in the planning mechanism. Based on this, the search tree node sort, the sampling node expansion, the target deflection expansion and the track evaluation and decision can be carried out, and the performance of the planning method under the uncertainty can be ensured. In the end, the simulation experiment under various working conditions is carried out. The single-step comparison experiment shows that the more accurate risk assessment can be realized by considering the trajectory planning of the uncertainty, so that the safer track can be determined; and the dynamic obstacle avoidance experiment shows that based on the accurate risk assessment, In the planning process, the search tree is expanded towards a more secure area, and the blindness and randomness of the planning under the uncertainty are effectively eliminated. Based on the key projects of the National Natural Science Foundation of China, a real-vehicle experimental platform is set up to verify the contents and methods of this paper. First, a design of a path following controller based on a single-point pre-scan and a research effort based on the speed of the pre-scan acceleration follow the design of the controller are performed. Secondly, in order to verify the real-vehicle experiment, the overall scheme design of a car as a real-vehicle platform is completed, including the design of the software and hardware architecture of the platform, the mechanical, the communication and the design of the power supply system, the final modification and the completion of the construction of the real-vehicle platform. Then, according to the real-vehicle experimental requirements, the platform parameters involved in the trajectory planning and track following control algorithm are estimated. Finally, the method involved in this paper is verified by the built platform. The experimental results show that the method can effectively improve the accuracy of the intelligent vehicle trajectory planning, and further ensure the running safety of the intelligent vehicle, and verify the effectiveness of the method described herein.
【学位授予单位】：吉林大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：U463.6

【相似文献】