Robot Localization Based On Spherical Panoramic Image

Artificial intelligence has developed rapidly over the last few decades due to rapid computer enhancements,and intelligent robots have begun to enter people's daily life.The first task to help mobile robots better serve human beings is to ensure robots know where they are,i.e.,robot localization.This thesis proposed an autonomous localization method based on a spherical panoramic image for a mobile robot following a specific walking path.The walking path was sampled at discrete points corresponding to identified locations.The spherical panoramic camera was installed on the mobile wheelchair robot,and data sets were collected at set location points for indoor and outdoor scenes.This study considered the localization problem as a classification problem,where each location corresponds to a category,and the approach's feasibility was verified using the popular Res Net50 classification network.Results confirmed that classification based on single location points can be easily confused between similar scenes.Therefore,we incorporated a long short term memory(LSTM)network to introduce time sequence information for robot positioning and hence improve localization performance.The major contributions from this study are as follows:(1)We transformed the complex robot positioning problem into a simple location point classification problem.Robot positioning can be solved as a classification problem since the spherical panoramic image not only provides 360° field of view,but also rotation invariance for each fixed sampling point,i.e.,image content is independent of camera rotation(robot direction);and we sample the route as discrete spatial points with unique labels,assuming that scenes captured at different sampling points differ from each other.Experimental results for the proposed method achieved high wheelchair robot localization accuracy using the spherical panoramic camera.(2)We introduced time sequence information for robot positioning based on classification tasks.Comparing experimental results for indoor and outdoor data,robot localization methods based on single images taken at single sampling points have some limitations,in particular that robots may become confused observing similar scenes.However,human beings tend to walk a few steps forward and backward arriving at a location to help identify their position through observation Therefore,we propose a solution for identifying location across similar scenes by combining images from adjacent locations to assist current location identification,i.e.,observing adjacent scenes to reduce location ambiguity due to visual similarity.Thus,this study proposes a convolutional neural network-long short term memory(CNN-LSTM)architecture to implement robot location.Experimental results confirm that the proposed method is very effective for robot precise positioning in similar scenes.