Shape-Based Segmentation of Objects in Cities using LIDAR Data

Abstract

This paper presents a system for recognizing and segmenting objects in 3D point clouds of urban environments within a probabilistic framework. Our system learns a probabilistic model of object shape from manually labeled training data. We then use this model and a boosting classifier to learn the relationships between recognition hypotheses (object location) and segmentation hypotheses (data points that belong to that object class). Finally, we conduct approximate inference to find the most likely labeling of test data by iteratively updating recognition and segmentation hypotheses until convergence. We evaluate the performance of our algorithm on the car and tree object classes using a truthed LIDAR dataset obtained in NYC, which contains approximately 900 cars and 700 trees.