Automatic Driving Target Detection And Direction Control Based On Convolutional Neural Network

Automated driving technology is the core of the control system is applied to vehicle AI technology,is the realization of "intelligent car","intelligent traffic" and "wisdom city" key,with automatic driving technology becoming more mature and the traditional auto industry increasingly saturated,self-driving cars will be based on artificial intelligence is the traditional automobile market and products of profound significance.Traditional automatic target detection of driving cannot meet the needs of present automatic driving,under the background of this article in view of the automatic driving deep learning target detection and direction control study,through the deep learning convolution neural network model to identify the target vehicle and pedestrian,judge the trend of movement,to predict the steering wheel rotation Angle.Compared with the traditional target detection method,it not only greatly improves the detection accuracy,but also solves the problem of uneven driver level,which is of far-reaching significance to the development of automatic driving field.First of all,the advantages and disadvantages of one-stage and two-stage are compared.One-stage YOLO series algorithm not only has good detection accuracy but also is extremely fast,which is very consistent with the requirements of target detection algorithm in the field of automatic driving.Therefore,YOLO series network is selected as the basic network.YOLO algorithm has some limitations in the field of automatic driving.In this paper,aiming at the insensitivity of YOLO algorithm to small targets in the automatic driving environment,a multiscale feature migration and fusion YOLOv3 network is improved on the basis of YOLOv3.A comparative analysis of the KITTI data set and the original YOLOv3 effectively improved the problem of small target recognition,and the experimental results of various target detection were improved.Considering that autonomous driving is usually used on mobile terminals,devices on mobile terminals are often constrained by hardware computing power.The YOLOv3 model was designed to be lightweight in order to realize autonomous driving at the mobile terminal with finite computing power.The yolov3-tiny model is used as the basic network training,and the KITTI data set is used as the training set of the automatic driving lightweight network model.The experimental results show that the detection effect of the lightweight model is not ideal.In view of the unsatisfactory effect of lightweight network model,a method to improve model accuracy by optimizing feedforward network is proposed based on the improvement idea in the previous chapter on the premise of guaranteeing real-time performance.Resnet18 network is adopted to replace the original feature extraction network,which improves the detection accuracy on the basis of guaranteeing real-time performance.Finally,the Unity simulator data set was built to train the automatic driving corner controller model.First,build a 3d virtual track on the Unity platform to make the virtual track close to the real world by adding different shadows,bright changes and shades.Because most of the straight lines in the actual traffic scene lead to unbalanced sample data of the virtual track,the method of category balanced sampling is adopted to solve the problem of unbalanced sample data.The image data with time information was input into the convolutional neural network autopilot model to predict steering wheel Angle.The predictive Angle is simulated by socketio server communication network model and simulator.