Learning the Behavior of Users in a Public Space through Video Tracking Yan, W. and Forsyth, D. "Learning the Behavior of Users in a Public Space through Video Tracking", in Proceedings of IEEE Workshop on Applications of Computer Vision (WACV) , 2005 Yan, W. and Kalay, Y.E. "Simulating the Behavior of Users in Built Environments", in Journal of Architectural and Planning Research (JAPR) 21:4, winter 2004.
Problem Statement Analyze mass data of human behavior in a public space Input: 8 hours of video in Sproul Plaza 3pm to 5pm for 4 days human observers to provide validation Output: statistical measurements that can be used to evaluate architecture design in terms of human behavior
The Tracking System Head detector Background model: averaging frames manually selected Intensity thresholding: assume dark head/upper body ROI Background Subtraction Intensity Thresholding Blob Merging
The Tracking System Tracking by data association Spatial proximity (sitting) and consistency in velocity (walking) Hungarian algorithm to link blobs from frame to frame a
Shadow Using geometric context to avoid the human blobs to be linked by cast shadows Compute the location of the feet Cut off the lower 2/3 of the blob
Results Counts Time of stay by the fountain 26 human 32 computer manual difficult On the 6m (10fps) dataset
Walking path Wondering people On the 6m (10fps) dataset
Large-scale results Without human evaluation Total number of people entered the plaza Total number of people who sat ~5% ~1% ~0.4%
Probability that a person chose to sit by the fountain depending on the number of people already sitting there.
Distribution of time of stay More Longer Secondary seating is more popular than primary seating
Walking path Wondering path
Simulations
Discussions Very clear problem statements Validate the system on a small data set before applying it to bigger ones