A final evaluation of pedestrian detection and tracking from a moving platform

Abstract

This work represents the fifth in a series of studies on safe operations of unmanned ground vehicles in the proximity of pedestrians. The U.S. Army Research Laboratory (ARL), the National Institute of Standards and Technology (NIST), and the Robotics Collaborative Technology Alliance (RCTA) conducted the study on the campus of NIST in Gaithersburg, MD in 2009, the final year of the RCTA. The experiment was to assess the performance of six RCTA algorithms to detect and track moving pedestrians from sensors mounted on a moving platform. Sensors include 2-D and 3-D LADAR, 2-D SICK, and stereovision. Algorithms reported only detected human tracks. NIST ground truth methodology was used to assess the algorithm-reported detections as to true positive, misclassification, or false positive as well as distance to first detection and elapsed tracking time. A NIST-developed viewer facilitated real-time data checking and subsequent analysis. Factors of the study include platform speed, pedestrian speed, and clutter density in the environment. Pedestrian motion was choreographed to ensure similar perspective from the platform regardless of experimental conditions. Pedestrians were upright in the principal study, but excursions examined group movement, nonlinear paths, occluded paths, and alternative postures. We will present the findings of this study and benchmark detection and tracking for subsequent robotic research in this program. We also address the impact of this work on pedestrian avoidance.