Deep Learning Based Loop Closure Research For Visual SLAM System

Synchronous Localization and Mapping(SLAM)are the key technologies of autonomous positioning of mobile robots in unknown environme nt.However,due to the cumulative error of the tracking algorithm,the pose calculation and the map construction can not be trusted after long distance travel.As an important part of SLAM,by analyzing the current sensor data to determine whether the rob ot is located in the area has been explored,loop closure detection solves the above problems.The current state-of-art visual-based loop closure detection systems use visual bag of visual words(Bo VW)model.These methods have achieved good results in the ideal operating environment,but in the actual outdoor environment,they are susceptible to a variety interference of environmental factors(illumination,weather,pedestrians,vehicles).In this paper,the popular convolution neural network(CNN)technique is introduced into the loop closure detection problem,to solve the above problem by using its great advantage in image semantic understanding.First,we built a data set of images to test loop closure detection system under a variety interference of environmental factors.Second,the feasibility of using the convolution neural network to extract and match the image to complete loop closure detection is demonstrated by using the data set above.Finally,a hybrid loop detection method based on CNN and Bo VW is proposed.The CNN model is used to filter the pedestrians and vehicles out of the image,and form an index of the image with its classification information,then classical Bo VW model is used for image feature extraction and matching in the loop closure detection.This method improves the anti-jamming ability of non-scene moving objects and the retrieval speed of the key frame image.In the view of above work,this paper gives a detailed description of the design and implementation of the test software,the experimental data are given and compared with the loop closed detection algorithm based on Bo VW model.The results show that the method is robust to pedestrian,vehicle and other environmental factors.