Project Mothra is an ongoing project in my research group to develop a video-based motion analysis system that requires subjects to wear a multi-colored jumpsuit in which each body segment is colored differently (Figure 1). There are many aspects to this project, but computer vision is definitely a large component, so I have chosen a subproblem of this research as my project for the class.
Figure 1. The multicolored suit in the reference position
The goal for my final project is to write a program that takes an image of a subject wearing the multi-colored suit while standing in the reference position (as shown in Figure 1), and segments the image into labeled ROI's corresponding to the differently colored body segments. The solution will use a segmented body model, but will not rely on manual initialization/labeling of the color spaces. Later, this will be used as the initialization frame in tracking the subject through various arbitrary motions.
I am not exactly sure which algorithm that I will end up using, but my research paper will look at a few different model-based image segmentation algorithms so that I can make an informed decision.
I will test the program on many different images, and determine its sensitivity/robustness to the camera position and orientation, as well as the effects of variations of limb positions and different subjects.
I will test the program on images of different subjects in the colored suit from different camera locations.
Worst: The model does not match up to the image, and segmentation does not work at all.
Best: The algorithm works for all subjects in any orientation towards the camera, and segmentation is perfect with no background noise or incorrectly identified regions.