Abstract: The aim of scene classification is to build a semantic context for various tasks of visual processing, especially for target recognition. Binocular vision system is now widely equipped with mobile intelligent robots. However, monocular images are currently mostly used for scene classification task. One can obtain lower classification performance by using features extracted from monocular image as the complexity of natural scene. In this paper, an approach based on binocular stereo vision for scene classification is developed. A feature descriptor of indoor scene is proposed, that is a vector extracted from planes fitting parameters in several specified regions. The scene is classified as open space and close space classes using feature extracted from disparity map with nearest neighbor method. Both open space and close space scene are classified into some subclasses using Gist and proposed feature descriptor. To test this approach, a dataset of four indoor scenes categories is created. The experiments show that this approach achieves excellent classification performance.

Key words: stereo vision, indoor scene, surround model, classification

摘要： 场景分类的目标是为各种视觉处理任务建立语义上下文，尤其是为目标识别。双目视觉系统现已广泛配备在智能机器人上，然而场景分类的任务大多只是使用单目图像。由于室内场景的复杂性，使用单目图像进行场景分类的性能很低。提出了一种基于双目视觉的室内场景分类方法，使用在一些特定区域里拟合出的若干平面的参数作为场景的特征。采用层级的分类方法，依据视差图，场景被分为开放场所类和封闭场所类，利用提出的场景特征和Gist特征对上述两类进行细分。为了验证提出的方法，建立了一个包含四种场景类别的图像数据集。实验结果表明提出的方法取得了较好的分类性能。

关键词: 立体视觉, 室内场景, 空间包络模型, 分类