SPiKeS-T


SPiKeS-T (SuperPixel-Keypoints Structure Tracker) is a model-free tracking algorithm that I developed during my Master's research at Polytechnique Montreal. In this approach, the target to be tracked is initially defined by a bounding box in the first frame of the video. The objective is to continuously track this target across all subsequent frames. My proposed method utilizes superpixels, keypoints, and a voting strategy to achieve accurate tracking. This work has been published in Machine Vision and Applications.

  • C++
  • OpenCV
  • CUDA
PDF source

Methodology

Model Representation

The target is represented by a Spikes-based model. A Spikes is a modified superpixel linked with surrounding keypoints to result in a structure easy to match.

Matches

Each spikes of the model looks for a match in the current frame. This video shows consistent matches as the red lines are parallel.

Votes

Each spikes votes for the target's center location. As each spikes is independent, even if some of them are hidden, the other ones can still vote to locate the target.