Research Of Visual Odometry Based On Deep Learning Algorithm

In recent years,intelligent robot technology has developed rapidly and has been applied in many fields.Some robots can complete tasks in situ,such as industrial robots on a production line,while some robots need to complete tasks in motion.Autonomous motion intelligent robot needs to acquire real-time position and posture(hereinafter referred to as P&P)change information to control its own motion trajectory.Visual odometry is the key technology for motion intelligent robot to obtain P&P information.The so-called visual odometry is to use the image frames obtained by the camera to estimate the P&P change information of the robot in motion,and it is a key structure module of the Visual Simultaneous Localization And Mapping(VSLAM)system.Regarding the visual odometry,the main research contents and results of this paper are as follows:(1)The methods of using image information in previous visual odometry algorithm can be divided into three types: characteristic point method,direct method and semi-direct method.When the application scene is an outdoor structured environment with unfavorable factors such as large illumination contrast difference,the P&P change information obtained by the above three methods is not accurate enough.In the above application scene,the image edge line feature information is relatively richer and more stable,and the line feature also contains spatial structure information.Therefore,the paper makes advantage of the image edge line feature information to carry out our visual odometry research.(2)Considering that the Convolutional Neural Network(CNN)has good capability of feature acquisition and feature learning,the paper constructs a CNN model to acquire the image edge line feature information.On this basis,in view of the fact that the deep learning algorithm can not only learn camera parameters,but also has good robustness to the illumination changes in the application scene,the paper constructs the deep Category-Aware Semantic Edge Detection Net(CASENet)model to acquire the image frame edge line feature information according to semantic classification.(3)Based on the time continuity of image frames P&P information and the spatial structure of line features,the paper constructs a coding model combining 3-Dimensional Convolutional Neural Network(3DCNN)with Long Short-Term Memory(LSTM)network,the image frame edge line feature information acquired in(2)according to semantic classification is coded by 3DCNN+LSTM,in order to acquire the high dimensional characteristics from the image frames edge line features and spatial structure information.On this basis,the paper constructs a model of visual positioning system based on Recurrent Neural Network(RNN)to realize P&P estimation based on edge line features.(4)The performance of the visual odometry based on deep learning algorithm proposed in the paper is tested by using KITTI data set.The test results show that in the outdoor structured environment,the visual odometry based on deep learning algorithm proposed in the paper still has strong robustness even if there are unfavorable factors such as large illumination contrast difference,and can meet the basic performance requirements of P&P estimation in VSLAM system.