Object tracking for neural network systems

This invention describes a system for tracking objects in videos by combining different methods. It uses a powerful but computationally intensive artificial intelligence (AI) system, called a trained network, to detect objects in only some of the video frames. For the frames where the AI isn't used, or to refine its results, the system employs two faster, simpler object trackers: one adjusts the AI's initial detections based on object movement, and another tracks objects in even more frames, potentially every single one, to maintain continuous tracking.