Research On Active Lane Changing Method Based On Multi-perception Fusion And Deep Learning Decision

Lane changing is one of the common driving behaviors,and the active lane changing is also an important research direction of intelligent driving vehicles.A good active lane changing function can improve traffic efficiency on the premise of ensuring driving safety;In addition,in order to avoid passenger discomfort,the autonomous lane changing behavior should also conform to the driving habits of human drivers as much as possible,Therefore,this paper conducts research on autonomous lane changing methods based on multi-source perceptual fusion and deep learning decision-making.On this premise,this paper first establishes a sensor fusion scheme for the intelligent driving vehicle to optimize the perception output and lay the foundation for the followup work;Then,the data-driven method is used to establish a lane changing decision model based on deep learning;Then the lane changing trajectory is planned based on polynomial curve;Finally,the model predictive control method is used to control the vehicle dynamics to complete the trajectory tracking.First of all,in the aspect of perception,the characteristics of millimeter wave radar,camera and other sensors commonly used at intelligent driving vehicles are analyzed,the functional targets of sensor fusion at the end of intelligent driving vehicles are defined,and the solution of sensor fusion at the end of intelligent driving vehicles is proposed,including data association,coordinate system transformation,target tracking and fusion.Firstly,the advantages and disadvantages of common data association methods are compared,and the Hungarian algorithm based on Markov distance is proposed as the data association scheme;Then introduce the relevant knowledge of coordinate system transformation;According to the characteristics of the sensor,the Kalman filter method is selected to fuse and track the sensor information;During the fusion process,the tracking,creation,numbering and other functions of the target are completed.Secondly,in the aspect of lane changing decision,a lane changing decision model based on deep learning is established.This part firstly analyzes the lane change process,considers that the lane change decision is a process of time series accumulation,and considers that the vehicle driving characteristic sequence can reflect the traffic conditions of the lane where it is located,so it adopts the Long Short-Term Memory neural network model that can correlate the time series characteristics;When handling the existing data,select the lane changing and lane keep samples based on the lane changing motivation to reduce the imbalance of the data as much as possible;The loss function,learning rate and optimization algorithm of the model are discussed.After comprehensive comparison,a lane changing decision model based on Long Short-Term Memory neural network is established;The selected data is used to train the lane changing decision model.The results show that the recognition accuracy of the lane changing decision and lane keep decision is more than 90%.Thirdly,in the aspect of trajectory planning and trajectory tracking control of lane changing,the trajectory planning model based on quintic polynomial curve and the control model based on model predictive control are established.First,a quintic polynomial trajectory cluster based on the preview distance is generated according to the driving conditions,and then an objective function integrating the efficiency and comfort of lane change is established.The optimal lane change trajectory is selected by optimizing the objective function;After that,the three-degree-of-freedom vehicle dynamics model is established and linearized,and the trajectory tracking model is established by combining the model predictive control algorithm.Finally,using the Carmaker/Simulink software platform,the sensor fusion program and the autonomous lane changing function are simulated and verified.Firstly,the simulation environment and sensor model are built in Carmaker,and the fusion program is built in Simulink,comparing the original output of the sensor with the data after fusion,it is proved that the fusion program can greatly reduce the impact of noise;Then the lane changing decision,trajectory planning and tracking control models are built in Simulink,and the vehicle dynamics model in Carmaker is controlled by the output control quantity to verify the integrity and feasibility of the autonomous lane change function.