Research On Visual Localization Algorithm Based On Semantics

With the development of the network and the popularity of wearable devices,people's demand for location information is increasing.Therefore,location-based services are gradually attracting researchers' extensive attention.As people spend about 80% of their time in the indoor environment,indoor positioning technology is a research hotspot in the current positioning technology.Indoor visual positioning technology,with its unique advantages of built-in sensors,is gradually replacing those indoor positioning systems required extra cost.In addition,the way of positioning by visual information is similar to the process by human eyes,which is more worthy of further study.In this paper,semantic segmentation and localization algorithm in machine learning are combined.Firstly,the research of visual localization technology and semantic information are studied,and the combination of machine learning and visual localization is analyzed.Secondly,this paper studies the application of semantic components in visual localization system.In addition,this paper has completed the following research on the problems existing in the offline and online phases of the visual positioning system:(1)Image retrieval stage of visual positioning system takes high time cost because of the large database,this paper proposes a fast image retrieval method based on the semantic and content,the method using machine learning divide image database into the semantic database,can effectively reduce the online retrieval time.Precision retrieval is carried out in the semantic sub-database,which ensures retrieval accuracy and saves time.(2)Aiming at the problem of low accuracy in large database,the SCBIR method is used to perform accurate retrieval by using semantic database.Aiming at the problem of high time cost in global feature extraction of the online stage,a feature point location method based on semantic constraint is proposed,which effectively reduces the location area and the feature point extracted in matching phase,in order to reduce time cost of the features extraction.In addition,due to semantic constraints,this method can eliminate a large number of mismatched feature points,and take into account the speed of feature extraction method and the accuracy of positioning method,so as to improve the performance of whole visual indoor positioning system.The simulation experiments above are presented.The results show that the proposed method can effectively reduce the time cost of retrieval in the online stage while ensuring the retrieval accuracy,and reduce the time cost for feature extraction.Finally,improve the performance of the whole positioning system.