Multi-Object On-Line Tracking as an Ill-Posed Problem: Ensemble Deep Learning at the Edge for Spatial Re-identification

Abstract

Online tracking is a feature of many state-of-the-art object trackers. When learning online, the data is limited, so the tracker learns a sketch of the object’s features. For a tracker to successfully re-identify the same object in the future frames in many different contexts, including occlusions, the tracker has to keep meta-data over time. In multi-objective inferences, this can exponentially increase the costs and is an ill-posed problem. This paper introduces a model-based framework that combines an ensemble of offline pre-trained models cascaded with domain-specific context for spatial tracking. Our method is efficient in re-identifying objects detected by any camera detector as there is minimal online computation. The second model uses a cosine similarity ranking of the label detected by the first model to find its corresponding set of raw images from the domain training set. A high score means model one has previously seen the object, and a low score amounts to a new detection. By using a two-stage AI-trained ensemble at the edge device, we show that the proposed tracker can perform 10 times faster with its precise detection, and the reidentification at the second stage is accurate, avoiding ID flipping for longer durations on video streams.