Research On Indoor Visual Positioning Technology Based On Residual Network Image Retrieval

With the increasing demand for Location Based Services(LBS)in indoor scenes,indoor positioning has played an important role in various applications.Because the traditional method was susceptible to interference from electromagnetic signals during the positioning process,and was highly dependent on infrastructure.Therefore,visual indoor positioning has become a research hotspot in the field of indoor positioning with its advantages of low cost,non-contact,and high accuracy.The key lies in the effective use of image key points,feature matching points and other information based on the epipolar geometry theory.When the existing positioning method or system based on image content and epipolar geometry was implemented,the positioning accuracy in the indoor environment was reduced due to the change of indoor light,the shift of key points,etc.,and the calculation of the user position was also caused by the poor robustness of the algorithm The increase in positioning error could not meet actual application requirements.To solve the above problems,this paper proposed an indoor vision positioning method based on residual network model image retrieval.First,it built an indoor image database for a small-scale and complex environment,and designed an image retrieval and positioning coordinate recognition system based on the adjusted residual network.Through the improvement of the convolution kernel,optimizer and principal component analysis in the network,the residual was improved.The feature learning effect of the network on noisy data or complex data.Secondly,through the cluster optimization of feature vectors,the extraction of depth features and the construction of visual vocabulary,the search matching accuracy and speed of user query images in the image retrieval stage were improved,thereby improving the real-time performance of user location calculations in the online stage.Finally,a basic matrix solution method based on interior point iteration was designed to avoid the large accumulation of errors in the calculation of pose information caused by improper selection of image matching points.The sensitivity to noise during the calculation of the essential matrix was reduced,and the robustness of the visual positioning system in position estimation was improved.Finally,the experimental simulation results showed that the visual positioning scheme proposed in this paper had better positioning stability and higher positioning accuracy in complex scenes with different indoor light and different key point displacement changes.