Research On Mobile Robot Relocalization With Deep Learing

With the increasing maturity of artificial intelligence technology,robots have gradually developed from industrial robots to intelligent mobile robots that are convenient for people's life.When the robot tracking fails or is powered on and restarted,the current pose of the robot must be restored by using the relocalization.Otherwise,the robot cannot continue to build the map or accurately locate the environment.Therefore,relocalization is of great significance for the mobile robot.The actual working scene of the robot is complex and changeable,such as the environment light changes violently or there may be dynamic objects in the scene,and the complex environment has a great impact on the traditional visual SLAM relocalization.Therefore,with the rapid development of deep learning,more and more scholars have applied deep learning to the field of visual SLAM.This paper mainly aims at improving the existing problems in the visual SLAM relocalization.The main work of this paper is as follows:1.Research on relocalization based on convolutional neural network.In order to realize the relocalization algorithm based on convolutional neural network,a convolutional neural regression network model suitable for global pose regression was established on the basis of convolutional neural classification network.Aiming at the phenomenon of over-fitting when the data set used in the network is small,a training strategy using transfer learning and stochastic gradient descent is designed to complete the relocalization task.2.Research on relocalization of long short-term memory based on video stream input.Considering that the robot can collect continuous images and long short-term memory can make use of the temporal correlation between data,a network model combining long short-term memory and convolutional neural network is established in this paper.In order to avoid the overparameter in manual adjustment of loss function,an uncertain loss function based on bayesian deep learning is proposed.3.Research on relocalization based on multi-task learning.In order to improve the relocalization accuracy of the model,a network model is established to realize the relocalization and visual odometry simultaneously.Moreover,the loss function of adding motion information is proposed for different subnetworks so as to realize the relocalization function with higher accuracy.4.Experimental research on relocalization based on deep learning.In order to verify the relocalization effect of the algorithm,the algorithm accuracy of the model was tested under outdoor and indoor data sets.Different exposure images were used to evaluate and verify the robustness of the algorithm for illumination.