Research On Fast Loop Closure Detection Based On Deep Learning For Visual SLAM

Simultaneous Localization and Mapping(SLAM)is the process of constructi ng an environmental map and autonomously localization for a movement device equipped with a specific sensor through the information acquired by the sensor.The Visual Simultaneous Localization and Mapping(vSLAM)system refers to a SLAM system using a camera as a main external sensor.In the long-term auton omous navigation operation of vSLAM,there will be error accumulation between adjacent frames,causing serious deviations in the optimization convergence of t he back end.The loop closure detection module in the visual SLAM system ena bles the mobile device to effectively identify the scenes that have been reached,thereby reducing the accumulation of errors between adjacent frames.A stable an d effective loop closure detection algorithm should meet the requirements of the visual SLAM system on accuracy and real-time performance.This paper has carr led out research work on these two aspects.1.Aiming at the problem that the accuracy of the traditional loop closure detection algorithm will be greatly reduced facing faster scene changes,illumination changes,weather changes and perspective changes,a loop closure detection algorithm based on pre-trained convolutional neural network is proposed.Firstly,the effects of different intermediate layer output image features using pre-trained convolutional neural networks are compared.Secondly,the image feature matrix is constructed based on image similarity strategy.Finally,using the cosine similarity,the newly added image frames are grouped in the image feature matrix to discriminate the loop closure.The experimental results show that the accuracy of the loop detection in the complex scene is significantly improved.2.Aiming at the high time consumption of traditional loop closure detection algorithm in the matching between the extracted image features and the constructed features description,a fast loop closure detection algorithm based on local sensitive hash is proposed.Firstly,utilized the theory of random hyperplane to constructed the local sensitive hash function.Next,the high-dimensional feature matrix is reduced and clustered by the local sensitive hash algorithm.Then,the newly added image frame is hash to obtain its hash index.Finally,the loop closure detection is performed inside the cluster based on the threshold.Comparative experiments on standard datasets show that the algorithm sacrifices a small percentage of accuracy in an acceptable range,but greatly improves the time performance of loop closure detection.