Research On Muti-way Coupling Intelligent Decision Oriented Towards Personification/Transhuman Driving

The difficulty of intelligent driving development at present is to develop the level of "personification",that is,how to make the intelligent vehicle make the decision of "personification" on the real road.In this process,a decision-making method based on rules and end-to-end is formed.In the end-to-end driving decision-making method,the vehicle uses neural network to learn the driving decision of intelligent vehicle through abundant driving data.This process is different from the rule-based decision-making method,which requires designing decision training network algorithm,collecting data to provide training scenes for the algorithm,and deploying the algorithm in real vehicle.Based on the above three points,this paper proposes a multi-directional coupling intelligent decision-making algorithm for humanoid / Superman driving,constructs the real scene driving data set,and builds a real vehicle verification platform for algorithm deployment.In view of the above contents,the specific work of the project mainly includes.(1)Aiming at the problem that the existing end-to-end driving decision-making "personification" is not enough to predict the single task,which leads to the lack of multidirectional coupling characteristics of intelligent driving,RGB images are collected,driving image spatial features are extracted by deep convolution network,and hard parameters are shared in convolution layer,and multiple fully connected neural network structures are constructed in the whole connection layer to predict the steering direction The multi task decision network(multi task)of driving and braking can realize the convergence of all driving tasks in the parameter optimization of shared layer of multiple driving tasks,which makes the optimal solution of model parameters have strong correlation and strong coupling characteristics of human driving decision,and the reliability of prediction is verified on the test set.(2)Because the end-to-end algorithm relies on human driving data,simulation data of simulator can not provide reliable training samples for real vehicle driving.The real data abroad focuses on the detection and identification of targets,and lacks the data of real driver end-to-end driving behavior.The subject built a prototype vehicle for data acquisition,equipped with visual sensors,and corner transmission Drive and brake opening sensor,through ROS,complete the communication of vehicle information,record human driving behavior,establish standard CSV format data set,and provide support for algorithm.(3)The real vehicle verification platform for algorithm deployment is built.The end-to-end driving behavior prediction is realized by ROS.The dynamic parameters of prototype vehicles are calculated.Servo motors are installed on steering wheel,brake and driving pedal respectively,and their peak and rated power are calculated;the highest and rated speed are respectively;the motor configuration is determined by peak and rated torque.Through CAN protocol and Autobox prototype communication,PID algorithm is used to control the real vehicle completion End to end driving.The experiment verifies the feasibility of the multi-directional coupling driving decision in this paper.It can realize the driving decision of lane keeping behavior and self vehicle correction without semantic segmentation and target detection.Innovations of the subject are as follows:(1)Multi task learning(multi task)using convolutional network transforms the multidirectional coupling characteristics of intelligent driving decision into the solution of optimal solution of multi task network with parameter sharing.It solves the shortage of "personification" of single prediction model.On this basis,the weight loss function of self-learning weight is introduced,and the weight of multi task is added to the optimization of network reverse propagation,and further the generalization of driving decision algorithm is improved.(2)The real driving data acquisition and verification platform is built,which makes the endto-end driving decision-making research not limited to the limited scene of the simulator,and provides real driving data for the research of algorithm and the development and test of intelligent vehicles.It fills in the lack of end-to-end real scene driving behavior data in the intelligent driving data set research.