This project demonstrates how inexpensive, accessible 3D sensors can provide accurate pedestrian tracking even in congested spaces. Multiple sensors can be linked to easily scale up to study larger spaces. With this kind of sensing, we can better understand how pedestrians navigate space, and optimize the hallways, sidewalks, and plazas of our cities for their safety and comfort.
As cities become more crowded and congested, we increasingly rely on cameras to study pedestrian movement in urban environments. Yet obtaining individual movement trajectories in space is often slow and expensive, requiring people to watch and manually annotate videos. Despite advances, computer image analysis remains inaccurate without some manual assistance.
We have developed a method of using Microsoft Kinect 3D sensors to perform reliable, inexpensive, and scalable pedestrian tracking.
At MIT, we suspended three Kinects from the ceiling over a busy hallway. Each Kinect collected depth images of people walking underneath in a rectangular area. Using hierarchical clustering, we automatically detected individual pedestrians with at least 94% accuracy, compared with manual annotation. The trajectories, accurate to a few centimeters, could also be stitched together across the three Kinect datasets, allowing us to identify the same walker as they passed from one camera to the next.