CS7322: Computer Vision II: Final Project Proposal

Human Gait Recognition

Robert Orr

The Idea:

In this project, I'm proposing three objectives:

1. recognize whether or not there is a person in an image sequence who is walking
2. differentiate between walking and running
3. recognize a person's identity based on gait features 

The Domain and the Scope:

Specify the domain and the scope of what you are trying out and why?  What is your data?

The questions I would like to answer in this project revolve around deciding which image features are most appropriate or effective to use in these recignition tasks. A number of different computer vision approaches have been develoeped for gait recognition, each with its own level of complexity, sophistication, and efficacy. It is my goal to choose a simple approach that is effective. These approaches include

• planar patch models of humans, and parameterization of patch model motion (Ju, Black, and Yacoob)
• spatiotemporal surface analysis (Niyogi and Adelson)
• cyclic motion analysis (Seitz and Dyer)
• multi-level recognition framework including HiddenMarkov Modeling (Bregler)

I plan to choose one approach and implement it. My goal is to build a system that can at least minimally perform my three stated objectives. I'd also like to learn about gait recognition along the way and have fun while doing it.

My data will be video sequences of different people walking or running. I will likely keep the number of different subjects to a minimum in order to simplify the problem and increase recognition accuracy.

Approach:

How do you plan to do this?  What methods?

I have five different papers on existing methods (see above). I will be reading each paper and will be evaluating each method as to effectiveness and ease of implementation. I am currently leaning towards the Niyogi and Adelson approach (XYT surfaces).

Proposed Time-line: