Research On Indoor Positioning Method Based On Computer Vision

With the rapid development of industries such as smart cities,Internet of Things,and mobile Internet,people's demand for high-precision indoor location services is becoming stronger.Traditional indoor positioning is mostly based on electromagnetic waves,sound waves,etc.,which belong to active positioning,which is easy to be interfered with and cause low positioning accuracy;while computer vision positioning is passive positioning and has stronger anti-interference ability.Predecessors used computer vision-based positioning work.For target positioning,most of them used artificially designed algorithms to extract image features,which made feature detection limited and unrobust;or compared the image of the positioning scene with the establishment of an indoor database for positioning.This method When the environment is changed,the database needs to be established again,and the positioning accuracy depends to a large extent on the perfection of the database.Computer vision positioning can obtain a wealth of environmental information,combined with current artificial intelligence technology,can achieve specific target positioning in a complex environment.This article starts from the combination of artificial intelligence and computer vision and their respective characteristics,and adopts the positioning method of eyes out.First,the camera imaging optical path is modeled,and the mathematical mapping relationship from 3D real space coordinates to 2D pixel space coordinates is described.Zhang Zhengyou calibration method and BP neural network are used to calculate the internal and external parameters of the camera.Secondly,it studies the target detection algorithm based on convolutional neural network,analyzes the role of convolution operation in target detection,and proposes to add a hollow convolution pyramid module to the YOLOv3 model structure to strengthen the model's small target detection ability;analysis The loss function of YOLOv3,optimized and improved by Focal Loss and CIOU,makes the model pay more attention to the training objectives of this article,and strengthens the effect of model training.In order to enable the method to achieve real-time positioning,this article optimizes the long-consuming target detection module,and uses Mobile Netv1 based on separable convolution as the feature extraction network of YOLOv3,and only uses a two-scale detection structure.The experimental results are obvious.The detection speed is increased,and real-time detection can be achieved.Finally,this article takes the robot as the positioning object and builds the experimental site,and completes the visual positioning experiment under the ROS framework.The 2D coordinates of the target detected by YOLOv3 are calculated in three dimensions,and the two-dimensional world coordinates of the robot in the experimental scene are obtained.Experimental results show that the method designed in this paper can achieve real-time positioning,and has good positioning accuracy,which meets the design requirements.At the end of the article,it analyzes and summarizes the shortcomings of this method and needs to be improved,and discusses future research directions.