Research On Indoor Place Recognition Based On Clustering Methods

Recognizing and understanding the indoor environment,which is involved with autonomous navigation and excuting semantic tasks,is very important for indoor serve robots.Indoor place recognition is the main method for indoor environment understanding,and it helps the robot to extract special semantic information and constructure the semantic map,which connect the space and semanteme.Also,indoor place recognition helps to solve the problem of relocalization for navigation and optimization of geometry map.This research is going to discuss the problem of indoor place recognition based on clustering methods,and improve the precision,speed and constancy for indoor place recognition.Firstly,a static indoor place dataset and a dynamic one are sampled,which are used to test static indoor place recognition and dynamic indoor place recognition algorithms.These datasets take many complicated factors into consideration,including illumination,occlusion,scale variance,visual angle variance and human factors.Secondly,a novel indoor place recognition algorithm CFI,which is based on feature clustering and image clustering,is proposed in this work.CFI extracts information by supervised feature clustering,during which global independent step and local independent step are the key parts.The recognition process is the clustering of images,where image similarity and corresponding clustering method are put forward.Besides,there is the state inertia to optimize the final result.The experimental results show that CFI has high recognition rate and speed for standard indoor place environment.Thirdly,in order to furtherly improve recognition speed of CFI,a new fast feature match algorithm based on r-nearest k-means is proposed.r-nearest k-means is a novel approximate nearest neighbor algorithm proposed in this work,it implement k-means clustering to the feature dataset first,then it finds the nearest centers to the query point and search the nearest neighbor in the r nearest clusters.The core ideal of this algorithm is substituting the global minimum value with the local minimum value.Experiments results show that this algorithm can speed it up over linear search by over 10 times with matching precision of 98%,and when CFI is optimized by r-nearest k-means,CFI's recognition speed is speed up by three folds.Finally,in order to solve the problem of that when time flies,the environment changed and the initial model cannot recognize the changing environment as good as former which reduse recognition decrease,a dynamic indoor place recognition algorithm is proposed.The dynamic algorithm is based on CFI and r-nearest k-means,which is a unsupervised learning process.It update the environment model by replacing the point that has high probility to be a goog recognition with a new point.The update process is controlled by confidence level and learning rate proposed in this work.The initial model will take in the information of the changing environment by unsupervised learning.Experimental results show that this dynamic algorithm keep high recognition speed for changing environment with remarkable speed.