Research On Autonomous Driving Method Based On Computer Vision And Deep Learning

Autonomous driving means the vehicle change the driving behavior in real time and complete the driving task by observing the surroundings without human intervention.Autonomous driving can reduce the number of traffic accidents,improve the road utilization rate of traffic resource and save travel cost.So the research of automatic driving technology is of great significance.Autonomous driving technology based on computer vision uses visual sensors as input and drive actions as output.There are three main typical methods as follows: Mediated Perception,Direct Perception and End-to-End Control.Among them,Mediated Perception method separates driving task into some sub tasks,including object detection,tracking,semantic segmentation,camera model and calibration,3D reconstruction and so on.Direct Perception method learns the key indicators of traffic situation and then be controlled by the control logic.End-to-End Control method establishes the input to the action mapping directly with a concise system structure.We design an automated driving algorithm based on End-to-End Control and deep learning method,which takes the problem of autonomous driving as a whole to study and establishes an end-to-end learning system.The learning system is a convolution neural network(CNN)consisting of seven convolution layers and four fully connected layers.The input of the network is the image from the first view of vehicle,and the output is a floating point number which represents the steering angle to be predicted.Compared to traditional methods which can only predicts movements like left/right turning etc.,the continuous steering angle is a more accurate description of motion.In addition,in order to improve the training effect,we use the network pre-training method and the overfitting prevention measures.Compared to the Mediated Perception method and the Direct Perception method,the proposed algorithm in this paper has obvious advantages.Firstly,the algorithm avoids the complex system structure of the indirect mapping system,because the Mediated Perception structure needs to be divided into several sub-tasks,and its design and implementation is very difficult.Secondly,the algorithm can accomplish data collection and network training in real situation efficiently.However,the Direct Perception method needs to learn the key indicators of traffic situation like the distance from the barrier,the distance from the marker line and so on.But the precise collection of these data in real situations often requires equipment such as ultrasound and laser radar,which are costly and difficult to implement.The proposed algorithm only needs to record the current field image and the turning angle as training set.In testing period,only the images from the first view of vehicle are collected,the continuous control of vehicle is implemented according to the steering angle predicted by the convolutional neural network.In order to reduce the cost of experiment with a real vehicle,a mini intelligent vehicle system was designed to be used for data collection and algorithm verification.The intelligent vehicle performs data acquisition,network training and testing in a self-designed environment with marked lines and obstacles.We also find that the convolution neural network can extract useful features for driving by itself.The result shows that the intelligent car was be able to planning a reasonable route to avoid obstacle in advance and maintain high auto-pilot rates in conventional scenarios.