Predicting the Movement of Occluded Objects Using Motion Models and Statistical Tracking
Date
2024-03-14
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Long-term occlusions pose a significant challenge to tracking systems, resulting in broken tracks which obfuscate results and lead to ambiguity when analysing a scene. My proposed solution to this problem, which I call the Occluded Trajectory Modelling (OTM) system, predicts the movement of occluded objects with a statistically-learned model of their movements while occluded, allowing for seamless tracking across occlusions. The model detailing the movement of occluded objects is extracted from unannotated data using Network Tomography and an objective function which extracts object occlusion times. I create a hybrid system, consisting of the OTM working in conjunction with an adapted Multiple Hypothesis Tracker, that is able to seamlessly transition between tracking visible objects and predicting their movement while they are occluded. This is done in real time without the need for visual features. I test my system on two datasets: the first consisting of drone footage with multiple environmental occlusions, and the second containing scenes with multiple non-overlapping cameras. The experimental results obtained from both of these datasets show that my system outperforms existing visual tracking systems in terms of correctly re-identifying objects after occlusion, without suffering a performance trade-off.
Description
Keywords
Occluded Trajectory Modelling, Occlusion Handling, Object Tracking, Movement Prediction, Multi-Camera Tracking
Citation
Grond, M. M. (2024). Predicting the movement of occluded objects using motion models and statistical tracking (Doctoral thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca.