Research On Visual SLAM Place Recognition Base On Hash Retrieval And Landmark Screening

Place recognition is used to determine whether the mobile robot has reached the scene that has been reached before,and it is often used in the optimization of visual SLAM loop closing.With the wide application of vision sensors in mobile robots,vision place recognition in complex scenes is a long-term unsolved challenging problem.The challenges mainly come from two aspects: first,the appearance of the image changes due to drastic changes in light,season,weather,etc.;second,the perspective of the image changes due to changes in camera shooting angle.Based on the description of landmark of convolutional neural network to match the landmark of images,it has achieved some results.However,due to the low efficiency of landmarks retrieval mechanism and the large number of dynamic information landmark in the image as the description of the image,the real-time performance of mismatch recognition is poor.According to this,this paper uses Mutip Cp LSH algorithm and dynamic landmark screening to improve the visual place recognition algorithm based on convolutional neural network,The details of this paper are as follows:(1)This work introduces a visual place recognition technology based on CNN,and analyzes its advantages and disadvantages in detail.(2)A hash based visual place recognition technology is proposed.A Mutip Cp LSH algorithm is used to replace the bidirectional nearest neighbor search,and the visual place recognition system is divided into offline and online parts.The hash table and query table are constructed offline to store the landmark features in the reference image set,and the global optimal landmark in the hash table and query table is retrieved online.Finally,the performance of the algorithm is evaluated by comparing three challenging datasets.(3)A dynamic landmark screening based visual place recognition technology is proposed,which introduces dynamic landmark screening to solve the problem of landmark image mismatch caused by dynamic information in the image.The Edge Boxes are used to extract the original landmark in the image,and the Faster RCNN is used to extract the dynamic landmark in the image.Finally,the idea of cross overlap is used to screen the landmark with dynamic information in the original landmark.