Research On End-to-end Human-like Decision-making Algorithm For Intelligent Driving

As the core key technology of the intelligent driving system,control decision-making represents the intelligence level of the intelligent driving vehicle and determines the safety and comfort of the intelligent driving vehicle.The control decision-making of intelligent driving vehicles mainly includes two methods: rule-based and end-to-end based.The end-to-end intelligent driving decision-making method can directly map the perception signal of the vehicle to the specific control parameters of the intelligent vehicle,thereby simplifying the decision-making of the intelligent driving vehicle.At the same time,with the "generalization" ability of deep learning,the "anthropomorphic" and "intelligent" level of intelligent driving decision-making can be improved.In view of the problems that the existing end-to-end intelligent driving decision-making algorithms have poor coordination between vertical and horizontal decision-making tasks which has risks of task imbalance,insufficient simulation of human driving behavior and thinking characteristics,insufficient interpretability,and strong dependence on sample data,this paper carried out research on the end-to-end human-like decision-making algorithm that integrates human driving behavior and experience,and realized the vertical and horizontal control decision of the vehicle.The main research contents of the thesis are as follows:(1)A method of vertical and horizontal adaptive equalization learning was proposed.Aiming at the problem of poor coordination between vertical and horizontal decision-making in the existing end-to-end driving decision-making algorithms,a multi-task learning algorithm and a dynamic weighting method are introduced and improved,by recording the respective loss functions of the vertical and horizontal tasks in the learning process,the relationship between the magnitude of the loss function and the learning speed is calculated,and then the weights of the vertical and horizontal tasks are adaptively adjusted;At the same time,in order to measure the balance of vertical and horizontal control decision-making task learning,an evaluation method that can quantitatively evaluate the accuracy and balance of vertical and horizontal control decision-making was proposed.Finally,relying on the Comma2k19 data set to conduct sufficient verification,it showed that the constructed adaptive equilibrium learning method can simultaneously ensure the accuracy and balance of vertical and horizontal control parameters in the decision-making process,thereby improving the safety and comfort of control decision-making.And the practicality is higher than the method of manually adjusting the weights.(2)An end-to-end decision network that fuses human driving behavior was proposed.Aiming at the problem that the existing end-to-end intelligent driving network is insufficient for the simulation of human driving behavior and thinking characteristics,starting from the characteristics of human driving behavior,an end-to-end human-like driving control decision including spatiotemporal characteristics,historical state characteristics and future characteristics was designed.Multi-layer convolution and long and short-term convolutional temporal memory network(Conv-LSTM)are used to extract spatiotemporal features from the visual perception image time series of the road ahead,at the same time,one-dimensional convolution and long-term short-term temporal memory network(LSTM)are used to extract the historical state features of the vehicle state information time series,and then the multi-task parameter sharing method is used to make control decisions on the steering wheel angle and vehicle speed of the current time and future sequences.Using the future sequence as an auxiliary task to supervise the learning of the main task at the current moment.Finally,relying on the Comma2k19 data set,the network performance was fully verified,showing good feasibility and superiority.(3)An end-to-end decision network that incorporates human driving experience was proposed.In order to improve the problems of insufficient interpretability and strong dependence on sample data in existing end-to-end intelligent driving decision-making methods,two driving experiences implicit in the decision-making of vertical and horizontal control parameters implicit in human driving vehicles(The corner determines the steering angle,and the obstacle distance determines the vehicle speed)were analyzed and summarized,and knowledge distillation technology is introduced.By adding two new networks that focus on learning curve curvature parameters and front obstacle distance parameters respectively,the driving experience of assisted longitudinal and lateral control parameter learning is integrated into the end-to-end intelligent driving decision-making network.Through the verification of the Comma2k19 dataset,It showed that the constructed end-to-end intelligent driving network integrating driving experience has better prediction accuracy,thereby increasing the interpretability and anthropomorphic level of the network to a certain extent.(4)By building an end-to-end intelligent driving data collection and testing system,the effectiveness of the end-to-end human-like decision-making algorithm for intelligent driving was tested and verified.First,based on the prototype experimental vehicle,equipped with vision sensors,corner sensors and inertial navigation systems,the vehicle data communication is established through ROS,and an intelligent driving data acquisition and testing system in a real campus scenario is constructed.The collection of real driving data in the real vehicle campus,offline training,and online testing in campus scenarios show that,without environmental perception such as semantic segmentation and target detection,this end-to-end intelligent driving decision algorithm could achieve lane keeping and deceleration and obstacle avoidance.function,which verified the feasibility of the proposed end-to-end intelligent driving decision algorithm fused with driving experience.