Online Algorithms for Classification of Urban Objects in 3D Point Clouds

The current technology in stationary laser range-scanning enables high-resolution acquisition of 3D data in a sequential fashion. Traditionally, range scans are processed offline after acquisition, which significantly slows down the procedure. In this work we alleviate this limitation by developing low-complexity, online detection and classification algorithms. These algorithms are innovative in that they classify points into 5 distinct classes (vegetation, vertical, horizontal, car and curb regions) and robustly determine the level of the ground without requiring any prior training or parameter estimation. To construct these algorithms we extract cleverly chosen summary statistics which significantly reduce the dimensionality of the data. This reduction enables us to contrast the different classes by appropriately chosen Markov models and then use online techniques to detect a transition from one Markov model to the other. The identification of the ground level is further achieved by taking advantage of statistical properties of the distribution of the summary statistics. Our algorithms also use contextual cues to verify the existence of specific classes of objects. All our algorithms take advantage of the sequential nature of data acquisition by running in parallel and labeling points on-the-fly. Thus, these algorithms can be potentially integrated with the scanner's hardware and provide the foundation for the construction of high-resolution 3D data scanners that classify data as acquired. We have run experiments using complex urban range scans and have evaluated the classification accuracy against ground-truth.