Research On Localization Of Indoor Service Robots Based On Machine Vision

Indoor localization is one of the key aspects for the study of indoor service robots,since it is the premise for accomplishing tasks like robot's navigation and route planning.Therefore,it is a hot topic for researchers both abroad and at home.Due to the obstacles and the flow of people indoors,the outdoor localization technologies such as GPS can not be effectively applied into indoor scenarios.As the robot vision technology has excellent qualities like low cost,light weight,little energy consumption and unlimited environment,as well as richness of information and features contained in images,and the sound stability,it is more suitable for a system to achieve precise positioning,and thus attracts lots of attention in the field of indoor localization of service robot.Since the indoor positioning technologies for service robots are highly challenging,at present the algorithms used for image matching are mostly traditional ones,but they have problems in terms of stability,generalization capability,and precision rate.This thesis will conduct a research by constructing an indoor location map and realize indoor positioning through robot vision,in order to achieve precise and fast positioning.The research includes the following aspects:Firstly,this thesis introduces the background and significance of the research,analyzes the current level of technology in this field,the strengths and weaknesses of some positioning methods,highlights the difficulties that needs to be overcome,and explain the criteria for indoor localization methods.Then,the thesis then goes on to introduce the theories of constructing an indoor localization map.To improve positioning precision,the thesis proposes a method of indoor mapping based on raster map.The method divides the indoor place into small raster with the same size.Then robots take images of the indoor scenarios,which forms a dataset containing the images of each raster and the position information of robots.The mapping effectively improve positioning precision and realtime performance,and more importantly,estimate the direction of service robots.Thirdly,we set up a localization method based on the raster map.To avoid the complexity of feature extraction and manual setting,and improve stability,real-time performance and generalization capabilities,we use the powerful self-training ability of convolutional neural network for feature extraction,and introduce an positioning algorithm of image matching based on convolutional neural network.We also propose a dropout strategy and cross-entropy loss function to improve LeNet-5 model to deal with the relatively low precision and poor real-time performance.The improved model is named FLeNet-5 which achieves a precision rate of 96%.Lastly,we conduct an experiment by setting up a dataset of images for an indoor mall,run the FLeNet-5 algorithm on Python and TensorFlow with accelerated GPU to realize positioning.The results show that the accuracy rate reaches 90%,and positioning precision is within 1.5 metres of diameter.The algorithm can achieve sound robustness,and relatively good generalization capabilities.