Research On Visual Place Recognition Technology Based On Convolution Neural Network Feature

Place recognition is based on data association and matching technology to determine whether the current location is where the mobile robot has been.Place recognition technology is often used in SLAM(Simultaneous Localization and Mapping)to realize the real-time localization of the robot,or to realize the global correction of the map by loop closure detection.With the wide use of visual sensors in the field of mobile robots,visual place recognition has become the key technology for robot to complete the autonomous navigation task,especially its application in a long-term and large-scale scene,which has become the basis for the rapid development of intelligent vehicles and other hot research fields.Because of the intense environmental changes in a long-term and complex scene,such as light,weather,season changes,the difficulty of the feature extraction of visual data is increased,and the feature description method of manual design such as SIFT is difficult to meet the requirement of place description in the visual place recognition task under this situation.CNN(Convolutional Neural Network),which has full and effective pre-training,has a very strong feature extraction capability for visual data,which can adapt to these environmental changes in visual place recognition.In view of the above,a new method based on CNN to extract image features and to realize place recognition by feature fusion,binaryzation and sequence image matching is proposed in this paper.First,a method of extracting image features using CNN model is proposed.On the basis of the comprehensive evaluation of the performance of the classic CNN network in the scene recognition task,a CNN model based on VGGNet is designed.By modifying the network structure and the layer parameters,the data of three convolutional layers are fused,and the data reduction is realized through the pooling layers,and the feature extraction and fusion are realized.Secondly,a feature binaryzation method is proposed.The feature data extracted from the CNN network can be converted from floating point data to binary data through a certain calculation method to achieve the feature's Hamming distance matching and improve the efficiency of the algorithm.Thirdly,a place recognition method based on sequential image matching is proposed.The image sequence contains the context information of the visual data,which can effectively eliminate the false positive results in the algorithm and improve the accuracy of the algorithm.In order to verify the effectiveness of the visual place recognition method,a comparative experiment with the current mainstream algorithms in challenging and representative open source data sets is conducted.Firstly,the structure and layer parameters of the CNN model are selected by experiment,and the feature extraction method is perfected.Then the sequence length of the sequence image matching method is worth setting by the experimental analysis.Most importantly,we compare the methods of this paper with the current mainstream algorithms FabMap,ABLE-M and SeqSLAM,and prove the robustness of the visual place recognition algorithm based on the CNN feature in a long-term and large-scale scene.As the key technology of autonomous navigation for mobile robots,visual place recognition plays a very important role in real-time positioning and map global correction in the SLAM process.This paper realizes an effective visual place recognition method based on CNN feature extraction,combines Hamming matching and sequence image matching,and hopes to play a role in promoting the development of robot technology.