Research On Decision Making System Of Autonomous Vehicle In Urban Environments

With the development of computer science and robot technology, autonomous vehicle is used a great deal in many ways such as military affairs, civil and science. The study of autonomous vehicle which has wide application prospects is a collection of the latest research results of structure, electronics, control theory and artificial intelligence.Intelligent decision is the key technology and hot area of research for autonomous vehicle. In urban environment, it is hard to model the behaviors of drivers, because that the driving scenarios are complex and the behaviors of the traffic participants are often unpredictable. For solving the problems of decision and planning for autonomous vehicle in urban environment, this paper presents a new establishing method of driving maneuver decision model by studying the decision process of drivers in complex scenarios, and then design the motion planner of autonomous vehicle based on the model. The research content in this article is listed below:1) This paper first introduce the research significance of autonomous vehicle and the international and domestic achievements, this includes the understanding the methods of robot intelligent decision and motion planning. Then the implementation of international autonomous vehicle is analyzed and compared, the traffic condition of urban environment is also described and summarized. This paper proposed the key problem of decision and planning for autonomous vehicle in urban environment based on the mission requirement, decision process and behavioral pattern of drivers, making the design criteria of decision making system clear. Then this paper introduces the composition modules of autonomous vehicle platform named "Intelligent Pioneer Ⅱ", describes the operating principle and collaborative ways of platform. The decision making system framework is consisted of three layered modules based on the multiresolution features in time and space. The design of decision making system meets the requirement of real-time performance, adaptability and robustness.2) For the different driving environment and driver’s maneuver features, this paper constructs the autonomous vehicle driving maneuver decision making module based on the theory of hierarchical finite state machine. After abstracting and decomposing the driver’s complex maneuvers, the atomic maneuver is used as the collections for states of state machine. At the same time, this paper presents a driving maneuver decision making model based on multiple attributes decision making method and the decision process of drivers in complex scenes. This model exacts the concerned relevant attributes of drivers, judging, evaluating and obtaining the final driving maneuver, making decision pattern more in line with the thinking process of drivers, solving the problem of humanoid decision for autonomous vehicle in urban complex traffic scenes. This paper designs a weighting method for driving maneuver decision matrix based on AHP and entropy weight method. After constructing the weight system based on driving experience and objective data, this method can recedes the interference with decision making results caused by subjective randomness and reducing the problem of inaccuracy of entropy method caused by the lack of sample data. This paper constructs a new driving maneuver gray ideal value model for decision making by combining the TOPSIS and gray correlation analysis method. This method makes the chosen alternative near to the ideal alternative both in position and shape, guaranteeing the optimality of the chosen driving maneuver.3) This paper studies the motion planning method based on Radial Basis Function Neural Network. At first, the design principles of motion planning algorithms in urban environment are analyzed, explicating the difficulty in motion planning. For solving the problems of that the road features are not clear and the environment is unpredictable, this paper presents the motion planning method based on Radial Basis Function (RBF) neural network. This method extracts the discrete reference point in drivable region. Regularization Network is used to approximate the data and Gradient Descent method which is a single output learning method with forgetting factor is used to train the RBF network parameters. Our approach produces the smooth, safe path and it can fit any road shape, meeting the requirement of vehicle kinetic characteristic. In addition, RBF neural network is a kind of local approximation neural network with the advantage of fast learning speed, so it can react fast to the dynamic environment and meet the real-time requirement of autonomous driving.At last, this paper performs experiments by taking "Intelligent Pioneer II" as the experimental platform in real urban environment. The results show the correctness and effectiveness of design methods of decision making system.