Evaluating Visuospatial Features for Tracking Hazards in Overhead UAS Imagery
Trevor M. Bajkowski, James Alex Hurt, Jeffrey J. Dale, David Huangal, James M. Keller, Grant J. Scott, Stanton R. Price · 2021
Tracking multiple objects as they enter and exit the field of view for a given image sequence has many applications in security, surveillance, traffic control, and additional important fields. To track objects of interest through videos a computer vision system needs two main components: an object detector, a model which can generate class predictions and bounding boxes for potential objects in a frame, and an object tracker, a model which partitions the nominated image regions (object detections) into groups containing the same real-world object instances. In this work we explore a method for object tracking involving clustering visuospatial features derived from an object detection model's outputs. For each frame, a variety of features are calculated based on the image patches nominated by an object detection model. Periodically, these detections and their associated features are clustered in each feature space. These feature-based partitions are then fused through evidence accumulation, resulting in a single partition of objects where each group corresponds to a single real-world object instance. To evaluate the aptitude of these features for the sake of object tracking, we test our approach on a set of low-altitude overhead UAS imagery for which object labels already exist. This work is meant to contribute to a larger system in which visual information is extracted from multiple airborne sensors in order to aid in mapping novel regions and the navigational hazards within.