General Online Object Deep (GOOD) Tracking Phase 2

Automatic high-value target trackers have a number of uses, including real-time sensor slewing on user-nominated targets, offline forensic investigations into where a specific target traveled, and assisting with automated missions running on remote platforms. Instead of automatically tracking all targets within a scene, it can be beneficial to focus on a single or small number of critical targets of interest in order to best utilize limited computational resources. In producing long, accurate object tracks there are a number of challenges, such as distractors that look similar to the target and the possibility of the target being occluded for unknown, unbounded periods of time. Kitware proposes to address these challenges through the General Online Object Deep (GOOD) Tracking System. At the core of this approach is a system of deep convolutional neural networks specialized for efficient aerial video tracking, coupled with advanced online learning to aid with long term track re-acquisitions. Re-acquisition is additionally aided by a scene model and periodic semantic segmentation applied to this model. The GOOD tracker will generate both highly accurate and long object tracks, which will greatly benefit its users.