Prof. Olympia Hadjiliadis (CUNY)

Title: Sequential on-line detection and classification in 3D Computer Vision

Abstract: In this talk we will address the problem of on-line detection and classification of objects in point clouds of urban scenes using sequential decision rules. We will begin this presentation by reviewing two well-known sequential rules in statistics, the sequential probability ratio (SPRT) and the cumulative sum (CUSUM). The former is used in sequential hypothesis testing and the latter in the problem of quickest detection of abrupt changes. We will then continue to describe the acquisition of 3D data points using laser range scanners and the limitations of the traditional off-line processing of this data, which creates a bottleneck. The goal of our work is to alleviate this bottleneck, by exploiting the sequential nature of the data acquisition process. In particular, we apply the above on-line algorithms and appropriate modifications of them to perform data classification on-the-fly as data is being acquired. We begin by introducing a cleverly chosen measurement model and then use appropriately tuned CUSUMs to distinguish vertical vs horizontal surfaces. We also present Hidden Markov models to capture vegetation in urban scenes. By applying CUSUM-like rules to detect changes from one Hidden Markov model to another we are able to identify the beginning of regions of vegetation. By then applying repeated SPRTs, we are able to identify the ending of these regions. We can thus distinguish vertical vs horizontal surfaces as well as regions of vegetation by making use of data sequentially. A far more challenging problem is the on-line classification of cars. One of the characteristics of cars is that they generate a large number of missing data (due to windows or metallic surfaces) and thus provide a big challenge to automated classification.

We use a combination of the above techniques, frequency analysis and a variety of measurement models, which integrate missing data in order to achieve their on-line detection and classification. Our solution does not require any training unlike most methods in the literature. Our algorithms can be potentially integrated with the scanner's hardware, rendering a sensor that not only acquires but also intelligently processes and classifies the scene points.

This is joint work with Drs. Ioannis Stamos and Hongzhong Zhang.