AN ADAPTIVE MULTIPLE-OBJECT TRACKING ARCHITECTURE FOR LONG-DURATION VIDEOS WITH VARIABLE TARGET DENSITY
Author ORCID Identifier
Date of Award
Department or Program
Multiple-Object Tracking (MOT) methods are used to detect targets in individual video frames, e.g., vehicles, people, and other objects, and then record each unique target’s path over time. Current state-of-the-art approaches are extremely complex because most rely on extracting and comparing visual features at every frame to track each object. These approaches are geared toward high-difficulty-tracking scenarios, e.g., crowded airports, and require expensive dedicated hardware, e.g., Graphics Processing Units. In hardware-constrained applications, researchers are turning to older, less complex MOT methods, which reveals a serious scalability issue within the state-of-the-art. Crowded environments are a niche application for MOT, i.e., there are far more residential areas than there are airports. Given complex approaches are not required for low-difficulty-tracking scenarios, i.e., video showing mainly isolated targets, there is an opportunity to utilize more efficient MOT methods for these environments. Nevertheless, little recent research has focused on developing more efficient MOT methods.
This thesis describes a novel MOT method, ClusterTracker, that is built to handle variable-difficulty-tracking environments an order of magnitude faster than the state-of-the-art. It achieves this by avoiding visual features and using quadratic-complexity algorithms instead of the cubic-complexity algorithms found in other trackers. ClusterTracker performs spatial clustering on object detections from short frame sequences, treats clusters as tracklets, and then connects successive tracklets with high bounding-box overlap to form tracks. With recorded video, parallel processing can be applied to several steps of ClusterTracker.
This thesis evaluates ClusterTracker’s baseline performance on several benchmark datasets, describes its intended operating environments, and identifies its weaknesses. Subsequent modifications patch these weaknesses while also addressing the scalability concerns of more complex MOT methods. The modified architecture uses clustering feedback to separate isolated targets from non-isolated targets, re-processing the latter with a more complex MOT method. Results show ClusterTracker is uniquely suited for such an approach and allows complex MOT methods to be applied to the challenging tracking situations for which they are intended.
Lohn-Jaramillo, Joachim, "AN ADAPTIVE MULTIPLE-OBJECT TRACKING ARCHITECTURE FOR LONG-DURATION VIDEOS WITH VARIABLE TARGET DENSITY" (2023). Dartmouth College Ph.D Dissertations. 119.