Research On Loop Closure Detection Algorithm Of Mobile Robot Based On Visual SLAM

Mobile robots are extremely dependent on the created scene map and localization in the map in unknown environments,and no matter what type of map is designed,it is inevitably affected by accumulated errors.And in the actual visual navigation of the robot,if the traditional loop closure detection method is used for scene recognition,it will encounter problems such as appearance changes caused by conditions such as time or season,changes in the viewing angle caused by sensors,and the target is moving violently.In this way,it is difficult for the system to extract the corresponding image features,and the extraction speed of image features is slow,which cannot meet the real-time requirements of the robot.Therefore,loop closure detection technology faces a lot of challenges and difficult problems.With the rapid development of deep convolutional neural networks,it is possible to enrich the image feature information extracted by visual sensors.In order to meet the accuracy and real-time requirements of loop closure detection deployed on the mobile platform,the following researches are carried out in this paper:(1)This paper proposes to combine deep learning technology with visual SLAM,and introduce a method based on convolutional neural network feature extraction in the task of visual loop closure detection.Aiming at the problem of poor image feature extraction accuracy of traditional algorithms,this paper designs the Shuffle Net V2-Net VLAD network model to extract features from images collected by visual sensors,output feature vectors,and then use cosine similarity calculation to perform similarity matching on image frames.Complete loop closure detection.Experiments show that the loop closure detection algorithm in this paper achieves higher accuracy than the traditional algorithm based on handcrafted features and the classical algorithm based on convolutional neural network.(2)This paper aims at the low real-time performance of traditional algorithms and deep learning methods applied to small and medium-sized devices.A loop closure detection method is proposed,which uses product quantization to reduce the dimension of the output vector of the Shuffle Net V2 network to obtain the compressed feature vector,and finally uses the asymmetric distance to calculate the similarity,which greatly reduces the high-dimensional spatial data operation time.The experiments have been compared with several excellent loop closure detection algorithms to prove that it has excellent performance in PR curve,average accuracy,time performance and similarity matrix,which can better meet the accuracy and real-time performance of the mobile platform.