Object segmentation and classification using 3-D range camera

This paper proposes a vision system using a 3-D range camera for scene segmentation and pedestrian classification. The system detects and segments objects in the foreground, measures their distances to the camera, and classifies them into pedestrians and non-pedestrian obstacles. Combining range and intensity images enables fast and accurate object segmentation, and provides useful navigation cues such as the range and type of nearby objects and the ground surface. In the proposed approach, a 3-D range image is segmented using histogram processing and mean-shift clustering. The ground surface is detected by estimating its normal vector in 3-D space. Fourier and GIST descriptors are then applied on each detected region to extract shape and texture features. Finally, support vector machines are used to classify objects; in this paper we focus on differentiating pedestrian and non-pedestrian regions. The performance of the proposed system is evaluated with two datasets. One dataset for object segmentation and pedestrian classification is acquired by us using a 3-D range camera; the other is a public RGB-D data-set for people detection. Experimental results show that the proposed system performs favorably compared to some existing segmentation and feature extraction approaches. Detecting and classifying objects in a 3-D scene plays an important role in assistive navigation for the blind [1], road safety [2,3], surveillance [4], and many other applications. In the traditional approach , visual recognition consists of image segmentation followed by classification [5]. Many image segmentation methods are based on low-level features such as color and texture. For example, Gould et al. proposed an object classification system based on multi-class image segmentation [6]. Their system labels pixels as background or foreground classes, and then classifies the foreground regions as cars, pedestrians or other. In an alternative approach, Leibe et al. suggested that the image segmentation and recognition are intertwined processes, and top-down knowledge from object recognition should guide the segmentation process [7]. Several top-down algorithms have been proposed to improve figure-ground segmentation of color images [7,5,8]. However, to avoid segmentation errors, several object detection algorithms without segmentation have been proposed, such as window scanning [9], local contour features [10], and implicit shape model [11]. With recent advances in 3-D cameras, range images have been used for object segmentation and recognition. Compared with color images, range images are less sensitive to changes in the environment illumination, object color or texture. Existing algorithms for range image segmentation focus mainly on segmenting …