The Design And Implementation Of Automatic Following System Of Intelligent Car Based On Convolutional Neural Network

With the development of robot technology,intelligent car plays a more and more important role in manufacturing,health care,entertainment and social interaction.Many scenes require intelligent car to follow its human leader.Traditional methods include ultrasonic positioning,UWB navigation positioning and so on.However,the range angle of ultrasonic positioning is limited.UWB navigation positioning requires that the car travel within the coverage of UWB base station.At the same time,both of them need to be equipped with target carrying device to complete the following.The intelligent car automatic follow system designed and implemented in this thesis does not need the target carrying device.It uses binocular vision technology to sense and obtain environmental information;processes and understands the information through the optimized convolution neural network to completes the positioning of the leader and to reduce the calculation force requirements;controls the car through PID control to complete the follow.The specific work is as follows:1)Using binocular vision camera as the only sensor and extract the depth information of the target from the left and right images based on the binocular camera model to complete the binocular distance measurement;2)Using convolutional neural network to find out the pedestrian area in the captured image.In this respect,based on densenet network,the thesis uses some ways to improve the feature recognition method like two-way dense layer technology to design DenseNet-Lite network;then optimize the SSD target detection algorithm to improve the speed by sacrificing part of the accuracy,and combine it with DenseNet-Lite network to make sure that high performance realtime object detection can be realized;3)Based on the results of object detection and binocular distance measurement,using Siamese network to compare the similarity and find out the specific coordinates of the leader.It uses PID control car to follow the existing leader coordinates.At the same time,the tracking strategy is optimized so that the car can continue to track and find the leader again when the leader is lost.In this thesis,the performance of each module is tested,and the test results show that the binocular ranging achieved in this paper has a high ranging accuracy for people in the short distance,with an average error of less than 0.05m;while the densenet Lite network combined with the optimized SSD achieves 70.9% m AP on the VOC 2007 data set,and can achieve a detection speed of 20 frames per second under the current hardware platform.At the same time,this thesis verifies the tracking system in different scenarios.The test results show that the scheme can complete the tracking task in various environments.After testing the tracking accuracy,the relative error calculated from the total error area root is 4.586%,which means the system can complete the following task accurately.