Research On Visual Location Based On Semantic Characteristics

In recently years,visual location as a kind of autonomous navigation method,has attracted much attention.This positioning method that uses visual images to determine its own pose has been widely used in indoor positioning,automatic driving,and robotics.At present,the basic process of common visual positioning mainly includes steps such as feature extraction,matching and motion estimation.Among them,the common feature extraction is to extract corner features of the image,such as SIFT feature points and ORB feature points.However,this pixel-level corner feature lacks the understanding of the scene,and it has become increasingly unable to meet the real needs of environmental perception and interaction with the complex environments.Semantic features as a more abstract feature can better achieve the surrounding environment expression.In addition,in the current common navigation maps,it is difficult to store the point cloud of feature points in the visual SLAM.Therefore,it is difficult for the visual location that based on corner point features to utilize common navigation maps.This paper proposes a visual positioning method that uses semantic features to determine the pose by extracting the semantic features in the image.The main work of this article includes the following three aspects:(1)The visual positioning method using semantic features is studied.Different from the traditional visual positioning based on local feature points,the introduction of semantic features can overcome some of the shortcomings of feature points,improve the ability to more abstract recognition of the surrounding environment,and enrich the diversity of image features in visual location.(2)Aiming at the problem of determining the position of semantic features in an image,a set of methods for extracting semantic features and restoring their geometric models are proposed and implemented.In order to achieve more efficient and accurate extraction of semantic features,this paper uses deep learning to detect the objects in the image.After training the required semantics through a large amount of data,the neural network can accurately detect the objects in the image.In addition,this paper combines multiple view geometry and point cloud processing methods to achieve the restoration of the semantic features in the image to the geometric model.(3)Aiming at the problem of matching of semantic features and positioning,an applicable matching and positioning method is proposed.Based on the semanticfeatures which have been determined location and the known map,this paper implements the matching and positioning method of semantic points based on the traditional point cloud matching idea.