Research Project P3:
Tracking People with an Overhead Camera

Research Project Goal

Given images from a stationary camera, find and track people. The video has been taken in a busy shopping mall in San Francisco. People enter and exit the scene in all directions. Sometimes, people walk as a cluster and occlude each other. The hope is that such cluster can be correctly segmented, and that people can be traced over long periods of time.

Tracking methods like this one will be useful in a number of applications. Some concern retail; at present, about a dozen companies market camera-based systems for characterizing the behavior of shoppers in retail. Others will be homeland security, such as the tracking of people in airports. At present we still lack methods for tracking people over longer periods of time using cameras.

Research Project Scope

Given video from a stationary camera of a busy environment (shopping mall, market place), track all people.

Data

These are two example images from two video stream containing several thousand images each. The instructors would be happy to provide additional video taken in different environments (hotel lobby, outdoor market).

Tasks

• Implement background subtraction method for identifying regions of likely change (homework assignment). Subtract background.
• Train feature detector for the shape of people. This detector will be dependent on where in the image we look
• Find possible people hypotheses by combining background subtraction method and trained feature detector
• Implement tracker for tacking people through multiple camera frames (particle filter? RANSAC?)
• Develop online learning methods for people's appearance (clothes, height, etc) to facilitate data association in tracking

Research Project Status

Team 1: Raylene Yung, Matt Jachowski, David Cohen

Team 2: Salomon Trujillo and Alex Perkins

Point of Contact

Sebastian Thrun

Midterm Report

not yet submitted

Final Report

not yet submitted