Face Tracking

Published

March 17, 2025

Problem Statement

This is a typical computer vision task. The goal is to improve the performance of the face tracking algorithm (see the video below). The algorithm is implemented in Python using OpenCV. Current version exploits thresholding segmentation and tracking based on simple features like energy and velocity vector.

Suggestions

One may consider other types of segmentation like edge-based segmentation. As for tracking, one may try a simple SORT algorithm using Kalman filter.

More ambitious ideas

Super resolution and face normalization

In ideal case, we would like to perform face normalization. However this is a very challenging task due to a very low resolution of our video data. We would need faces to be at least 60 x 90 pixels. To increase the resolution, we may consider using a super-resolution algorithm.1

Footnotes

  1. For example, Azure OpenAI can leverage models like GPT-4 and DALL-E to enhance the resolution of thermal images, making them clearer and more detailed.↩︎