Model Based Interpretation of 3-D Range Data

Abstract

One of the fundamental problems in computer vision is to recognize 3-D objects in a scene. This paper describes a model-based approach for recognizing 3-D objects by using 3-D range data. The vision system consists of four system modules: object modeling, model-driven prediction, feature to model matching, and 3-D feature extraction. Model objects are represented by a viewpoint-independent volumetric model based on generalized cylinders. The vision system first extracts global 3-D features from the range data to select possible object models and to estimate coarse object orientation. Three-dimensional physical edge features are then extracted and classified into occluding, convex and concave edges. These three types of physical edges correspond to multiple levels of object features. The occluding and concave edges are used for locating the contours of object components that can be modeled by individual generalized cylinders. The convex edges are used to locate surface boundaries within individual cylinders. Predictions of 3-D image features and their spatial relations are generated from object models to give guidance for goal-directed cylinder extraction. The results of feature extraction and model prediction are gathered at multiple levels and a reasoning process is applied to recognize the object. Interpretation proceeds by comparing the extracted 3-D features with object models in a coarse to fine hierarchy. Since the 3-D information is available in range data, the actual measurements are used for feature to model matching. Preliminary results on 3-D vehicle objects are shown and future research directions are discussed.