Design And Implementation Of Navigation System Based On Deep Learning

In recent years,due to its broad application prospects in logistics,environmental protection,traffic pressure relief and other fields,autonomous driving technology has attracted the attention of academia and industry.Because of its low cost and insensitivity to signal interference,visual navigation algorithm has become a hot research field.With the rapid iterative update of hardware and the continuous increase of computing power,the field of artificial intelligence changes with each passing day,and deep learning develops vigorously,and remarkable achievements have been made.Affected by this,many advanced algorithms have emerged in the field of computer vision,and breakthroughs have been made in the fields of pattern recognition and image segmentation.Using deep learning tools to solve problems in visual navigation is a promising research direction.We designs and implements a visual navigation system based on deep learning,which forms a complete solution from data collection and calibration to all-weather visual positioning.Among them,one of the key tasks is to design the visual navigation system and improve the algorithm.In our system,RatSLAM algorithm is firstly analyzed and deconstructed to solve the problem of data set annotation.Firstly,the initial map was generated for video frame,and then the map accuracy was corrected and optimized through feedback loop.Then the data structure of video frame and experience map is structured to get the picture after position calibration.In the location model part,DeepVO algorithm is improved by analyzing the optical flow network.Images are encoded by the optical flow network,and the information flow containing the optical flow information between images is obtained.The information flow is input into the two-layer LSTM network to output the pose information.Then,the pose information is integrated through an SE(3)layer,and the position is output.In the all-weather navigation part,we proposes a multi-domain style conversion model inspired by dual GAN and conditional GAN,which can convert evening or evening style pictures into daytime.Input the unified style image into the visual navigation model to complete the all-weather navigation task.In addition,the multi-domain style transformation model can not only solve the all-weather visual navigation task,but also has strong generalization ability,and has achieved good generation effect on other image style transformation tasks.We makes an overall experimental evaluation of the visual navigation system based on deep learning and verifies the reliability of the system from data collection to navigation results in the square in front of the main building of China campus of university of electronic science and technology.Experiments show that the whole system can complete the navigation task.