Research On Automatic Steering Technology Of Intelligent Vehicle Based On Deep Learning

Automatic driving is the key technology to realize "smart car","intelligent traffic" and "smart car networking",with the development of the traditional automotive industry gradually saturates,it will be the inevitable trend for the car to flourish in the future.With the rapid development of deep learning research,automatic driving has also come to a revolutionary age.This thesis studies the auto-steering technology based on deep learning.By applying deep learning to the sensing and decision-making processes in automatic driving and directly predicting steering wheel control commands to realize automatic steering of smart cars,the development of automatic driving technology will accelerate.This is of great significance to the final realization of automatic driving.Firstly,a research platform for autopilot system was set up.The modification of a pure electric vehicle was carried out,use the steering motor and the push rod motor to drive the steering wheel and the brake pedal respectively.For the throttle,an external DA signal can be applied from the electronic circuit to drive it.In the hardware aspect,the sensor platform and the controller computing platform were set up respectively.In the software aspect,a real car software platform based on ROS system and autopilot simulation software platform based on Prescan,Simulink and TensorFlow were build.Secondly,a smart car auto-steering model based on deep learning was established.The model was verified from two aspects of simulation and real vehicle test.The simulation results show that the model can converge well.Based on this,further training results based on actual vehicle test data show that after pre-training the encoder and training the steering wheel angle prediction model,the loss function converges faster,and the predicted angle can change well with the change of the actual angle.In order to analyze the road areas that the intelligent car auto-steering model focuses on,decoder roadmaps,back-propagation visualization,and deconvolution visualization methods are used to restore road images.The pictures show that under the lane keeping conditions,the model focuses on the edge of the road.Finally,the smart car auto-steering model was improved.The original steering controller network was replaced by long short-term memory network.The improved model used serialized image data as input to calculate steering wheel angle.The training results show that the improved model converges faster and the predicted angular quantity is more accurate and gentle.On this basis,the test in the intersection and lane change process was performed on the simulation software platform.Simulation results show that the improved model can predict the corresponding steering wheel angle according to the intersection arrows,and the lane change can be done based on the obstacle information in the current lane and the side lane.This has some implications for the realization of selective end-to-end deep learning.