Natural Scene Perception Based On Binocular Stereo Vision

Natural scene perception aims to make computer vision systems be capable of perceive scene as human being. The ability to rapidly and accurately perceive natural scene plays a vital role in intelligent image analysis system, which can narrow the search extent of objects, exclude the objects of inconsistent semantic interpretation in context .In last decades, natural scene perception has been a hot research topic in computer vision community with increasing demand to higher intelligence on images process that is necessary to intelligent system of vision.Local semantic concept model BOHOG2 and semantic properties model HOG+3gist are proposed in which improved local feature descriptor is used to counter lower accuracy when using gist feature to classify scene and expensive computation cost during gist feature extracting. BOHOG2 takes combined HOG and 3 scales gist as improved local feature descriptor of image, visual words are obtained by clustering local feature of images in training set which are used to quantify local feature, the content of scene is described by distribution of visual words among image. HOG+3gist directly use improved local feature to approximate semantic properties of scene. We trained SVM classifier on two natural scene image databases using features given by these two proposed model. The classification results show that the classification accuracy of two proposed model are higher than original gist feature. Additionally, we compared the time consumed during feature extraction by HOG+3gist and original gist feature, the experiments show that our proposed methods can speed up features extraction.As to several state-of-art approaches of scene perception perform poorly in most of indoor scene categories, a binocular stereo vision based indoor scene perception methods is proposed. Disparity maps of indoor scenes is modeled by a series of planes which are obtained by binocular stereo vision , scene is characterized as a set of normalized parameters by fitting planes using disparity map in appointed areas and proposed HOG+3gist feature. To test proposed approach, we trained SVM classifier on constructed stereo indoor scenes collections, mean classification accuracy achieves to 73.7% using proposed approach which is 13.2% higher than HOG+3gist feature. The experiment shows that combining 3D structure information of scene obtained by binocular stereo vision and features extracted from 2D image to describe scene can improve classification accuracy than only using features obtained from 2D image.