Research On Autonomous Navigation System Of Intelligent Vehicle Based On Deep Learning

For the shortages of recognition accuracy,real-time performance,single navigation system function,and anti-jamming ability of the core functions of intelligent vehicles’ visual perception,an improved Siamese network of object recognition algorithm is proposed to overcome above specialties in traditional image processing.Moreover,a road extraction algorithm for remote sensing image based on improved U-Net network is introduced to the navigation system,with optimizing the recognition ability of global path information,providing prior information for relevant route planning and navigation,enriching the diversity and reliability of navigation system.Finally,an embedded vehicle experimental platform is built to verify the proposed algorithms.The main research contents and results are as follows:(1)Analyzing and summarizing the developing technology of intelligent vehicle,including object recognition and tracking,intelligent navigation,as so on.Introducing the main research contents and objectives of this thesis.(2)Aiming at the function of intelligent video object recognition and navigating control system,contrastively studying the relevant deep learning networks.The different specialties of stacked autoencoder,convolution neural networks,and Siamese network of object tracking,the fully convolutional networks,deconvolution network,and U-Net network of road extraction are discussed.Comparing and analyzing the algorithm in combination with the application environment.(3)By introducing weighted channel and bounding-box regression strategy,a recognition and object tracking algorithm based on improved Siamese network is proposed to overcoming the errors caused by shadows,illumination and background changes frequently.The standard data set is adopted to train for tracking the object accurately.The experimental results show that the average IoU(Intersection over Union)is 62.19,which exceeds 58.40 of the former algorithm,the effectiveness of the improved algorithm is proved.(4)The dilated convolution and batch normalization are added to the original U-Net network,which realizes road feature recognition and extraction of remote sensing image.Compared with other algorithms,the results obtained in this thesis are more clear and complete,the average Dice’s coefficient is 0.74.It provides essential prior information for global path planning and behavior decision making.(5)An embedded vehicle experimental platform based on Raspberry PI 3B+ is founded,system integrates environmental perception unit,calculation and control unit,interaction and motion control unit.The improved Siamese-FC algorithm is used to the platform for object tracking.The feasibility of the algorithm is proved.The results show that the optimized Siamese-FC algorithm can locate the object accurately and efficiently,it is more adaptable to interference of shadow,illumination,and frequent background changing.The remote sensing image path extraction algorithm based on improved U-Net network is able to get clear driving path and provide prior information for navigation system.The real-time object tracking and autonomous navigation of the car platform are realized through the application and validation of the algorithm on the car experimental platform.