Overview

This was a project for CMU’s 16-720 B Intro to Computer Vision course. The task was to create an image tracker from scratch and apply several variants including, Lukas-Kanade with translation only, Lukas-Kanade with affine transformations, and the inverse compositional alignment algorithm.

Process

In order to implement the different trackers, I had to implement some linear interpolation in order for the tracking square to properly track the vehicle. Since the pixels are discrete and often does not express the level of granularity the LK algorithm expresses, interpolation is required to “add more pixels”.

Results

Firstly, here are the results from the inverse-compositional alignment tracker. This is a video, and I have selected six frames from that video.

• 

• 

• 

• 

• 

• 

The plane landing shown below implements the Lukas-Kanade tracker with translation only. This was a good video to use as the difference between frames was not extremely large.

• 

• 

• 

• 

• 

• 

This was a particular fast paced tracking video. It used the the Lukas-Kanade tracker with affine transformations. Notice how the bounding box expands as the frames progress. The tracker has issues with faster moving pixels.

• 

• 

• 

• 

•