Research On Key Technologies Of Weak Sensing Intelligent Parking System Adapting To Complex Environment

With the rapid development of environmental perception,automatic control,artificial intelligence and other high technologies,the intelligent vehicles have become the research focus and the new increasing-point of the automotive industry in the world.In the road map of the development of intelligent connected vehicle,intelligent parking technology is regarded as the best breakthrough point for the application of intelligent vehicle.Therefore,the research and development of intelligent parking technology has become one of the research focus in the global automotive engineering field.The application of intelligent parking system on many models exposes the application problems and technical difficulties.This also make the research direction of intelligent parking key technologies more concentrated and clearer.For this reason,this paper studies the intelligent parking technology in weak sensor environment.It includes the intelligent modeling method,the dynamic parking trajectory planning method and the control method of path tracking,parking performance comprehensive evaluation method and so on.The main work is as follows:Firstly,aiming at the inconvenience of Cartesian coordinate system used in most parking modeling and the positioning and attitude deviation caused by simplifying parking model.The frenet frame is introduced to dynamically record the scene map of database searching process.Based on deep learning method,the parking process is intelligently modeled.In order to improve the accuracy of the model,the dynamic effect of low speed motion is incorporated into the model.Finally,based on the kinematics model,the dynamic parameters in the process of parking movement are considered.The intelligent parking model is established by neural network method,which lays the foundation for the path tracking control.Secondly,in order to solve the problem of the excessive obstacle boundary deviation because of the weak sensor,we select some typical parking scenes.Then the experiment is carried out according to the parking requirements of IPA system.Meanwhile,the data such as vehicle speed,long-distance ultrasonic sensor ranging value,echo width,echo height and steering wheel angle are collected,and the training set,test set for deep learning are constructed.Then,according to the output characteristics of the data,the improved flow clustering algorithm is designed to distinguish the data sets corresponding to different obstacles from the continuous data.Next,a semi-supervised deep learning model for obstacle boundary recognition is proposed,which combines the high efficiency of supervised learning and the good adaptability of unsupervised learning.A semi-supervised self-coding obstacle boundary recognition method based on label and sparse regularization constraints is established.Eventually,based on this,the method of location recognition is proposed,and the algorithm is tested and verified by the test set,which lays the foundation for the path planning.Thirdly,aiming at the problem that the existing path planning methods focus on the vehicle,the parking space and its surrounding obstacles are only used as collision constraints.Its influence on the parking trajectory or tracking control is not fully considered.The concept of parking trajectory model based on particle interaction is proposed by referring to the relevant theories and methods of standard particle model in high energy physics.The intermolecular force model is introduced to describe the interaction process between the surrounding obstacles and the vehicle.Then we introduce the quark interaction model to describe the attraction relationship between the vehicle and the target location.Subsequently we builds a parking model based on particle interaction to describe the forces of the vehicle in the parking environment.Then,the parameters of the transfer function from the particle gravity and repulsion to the motion of the vehicle are determined by the optimization method.The input of the model is the distance between the vehicle and the obstacle,the difference between the vehicle and the target parking position,and the current position and posture of vehicle,etc...The output is the target position and posture of the vehicle in the next period(including the reference control information such as the target speed and the target steering angle).After the model is established,the performance of path planning is simulated in each typical parking scene to verify the generalization performance of the parking trajectory planning method based on standard particle model.Fourth,the vehicle shows strong nonlinear characteristics when parking in a low-speed,which aggravates the deviation of longitudinal and transverse control.This brings challenges to the tracking control of parking planning path.In this paper,the real-time posture error model of vehicle is designed based on the parking intelligent model established in Chapter2.The nonsmooth control theory is introduced to solve the nonlinear characteristics of vehicle.The nonsmooth control model of trajectory tracking is established by taking the realtime position and posture of vehicle,and the target position and posture as input,the error model output as feedback.Then a non smooth control strategy for anti disturbance tracking of parking track is proposed.The control command of the chassis control unit can be generated directly by taking steering angle and vehicle speed as output.Finally the tracking control of parking track with high accuracy and efficiency can be realized.Finally,this paper combined with domestic and foreign standards,technical requirements of enterprises and market research needs,proposes a parking performance evaluation system.This is for the complicated scene and disorder of the indicators in the parking performance evaluation.It includes subjective and objective evaluation indexes,which can be subdivided into seven specific indexes.Based on the theoretical and technical research results of this paper,an experimental system is built to carry out the real vehicle performance evaluation experiment.In view of the problem that the real vehicle experiment takes a long time and the similar scene experiment sample is insufficient.So it is difficult to obtain the convincing parking performance stability evaluation results.In this paper,the bootstrap method and appropriate real vehicle test sample are adopted.By fully applying the information of the sub sample to resampling,the small sample is proposed.The evaluation method of the performance stability of IPA System is carried out based on the real vehicle experimental data.To sum up,based on a large number of real vehicle scene experimental data,this paper proposes a data-driven method to establish parking intelligent model modeling method,and a parking slot recognition method based on deep learning.In order to eliminate the influence of the environment on trajectory planning and tracking control,this paper proposes an trajectory planning method based on standard particle model and an anti disturbance tracking control strategy based on nonsmooth control theory.In view of the difficulties in parking performance evaluation,this paper puts forward an evaluation index system by combining the subjective and objective factor.Then an evaluation method for parking performance stability is also given.It provides a theoretical basis for the research of intelligent parking technology under the condition of weak sensing.