Progressively Expanded Neural Network (PENnet) Tracker

We propose to develop a multi-layer tracking framework that integrates feature-based, motion-based, and deep-learning-based tracking. Our proposal consist of two development tasks and two research exploration efforts. Our team proposes to develop the Progressively Expanded Neural Network (PENnet) Tracker, a deep learning based tracker capable of on-the-go training of the target. The PENnet Tracker will retrain the deep network every few seconds and with each iteration the training set will get larger, thus allowing PENnet to better learn the target representation. Next, we propose integrate the PENnet Tracker into a multi-layer framework. Additional layers will include a feature-base and motion-based trackers. We propose using the robust Directional Ringlet Intensity Feature Transform (DRIFT) as our feature based tracker. We will conduct an exploratory effort to determine the optimal approach of using deep networks as a high-level decision maker. And lastly, we will conduct a study to determine the effect of frame-rates on tracking performance. We will report on the impact and optimal operating rates for various trackers.