Person detection : unmanned system and small sensor applications

Abstract

The ability to quickly and reliably detect people in images and video is highly desired. Several object recognition algorithms have demonstrated successful detection of multiclass objects with varied scale, position and orientation. This study examines the effectiveness of these methods when applied to detecting humans in two distinct domains: A) Leave-behind sensing and B) Aerial surveillance. Using novel image sets that are significantly more realistic and difficult than standard datasets, a variety of tests are conducted to compare the algorithms in terms of classification success rate. Dalal and Triggs' Histogram of Oriented Gradients algorithm, when trained with image samples taken from inside MIT's Stata Center, detects with no false positives all but one person in six minutes of video taken from inside a separate building. An enhanced version of Riesenhuber and Poggio's cortex-like recognition model, trained to detect people, correctly classifies 95% of images taken from a small UAV when trained with an independent set of images. These results illustrate the potential to accurately and reliably determine the presence of people in video from unmanned aircraft and indoor sensors.