Research On Recognition Technology Of Indoor Scene Based On Multiple Features Fusion

Scene recognition is an important research direction in the field of computer vision.The scene recognition of indoor mobile robots requires that the robot can use visual information to judge the category of the environment or scene.Such as living room,bedroom,corridor and so on.In this paper,the methods and processes involved in the scene recognition technology based on multi feature fusion are studied.Firstly,the global and local characteristics of the scene image are extracted,Then the two features are fused with the multi kernel learning method and the scene is classified.Finally,a large number of scene recognition experiments are carried out using the software developed by ourselves.The work of this paper is summarized as follows:Firstly,this paper adopts the feature extraction method combining the global and the local for the indoor scene.The H/I one-dimensional color model is established,It can quickly and accurately extract the global color information of the scene images;Then,a SURF-DS-BoW feature extraction algorithm based on SURF descriptor and Dense-SIFT descriptor is designed.SURF can describe the foreground area with significant changes in the scene,and Dense-SIFT is conducive to describe the background area that changes slowly.When generating BoW features,the two features can be weighted together to express the image characteristics of the scene more accurately.Secondly,In the process of mapping an feature vector to a visual word.Using the hierarchical adaptive K-means can effectively improve clustering accuracy and reduce the representation time of BoW.a new BoW model SWSA-BoW which is based on new weighting method named Similar Word Soft Assignment(SWSA)is proposed to abstract the features of local areas of indoor scene picture.The method of Similar Word Soft Assignment will assigns the extracted features to several visual words which are closer to each other and give different weights.It can effectively solve the effects of synonym and ambiguity between visual words,It also can improve the ability to distinguish visual words and overcome the shortcomings of traditional soft allocation methods,and then improve the accuracy of scene recognition.Thirdly,the global and local features are fused by multi kernel learning to improve the performance of the support vector machine(SVM)classifier.Firstly,the optimal kernel function of each characteristic corresponding to SVM is determined.Then multi kernel learning is used to optimize the weights corresponding to each function.Compared with single kernel,based on multi kernel learning support vector machine(SVM)can get better classification performance.Finally,this paper depending on Microsoft Visual 2012 and OpenCV software,the indoor scene experiment are completed.The experimental results show that the proposed method can recognize the scene efficiently.