Description
Presented our work entitled "Occlusion-Resilient Cattle Tracking in Barn Environments Using Monocular Depth Estimation and 3D Relational Bounding Boxes" at the 1st EAAP Conference on Artificial Intelligence 4 Animal Science (AI4AS 2025), held in Zurich, Switzerland as part of the "Advancing Digital Biomarkers with AI: Breakthroughs in animal identification, health and welfare monitoring, behavior analysis, and remote sensing technologies" session. Introduced a novel approach for improving cattle tracking in visually complex barn environments by combining monocular depth estimation with 3D relational reasoning. The method enhances tracking robustness under occlusion, supporting more accurate behaviour analysis and welfare monitoring in precision livestock farming.Associated Abstract:
Occlusion, or loss of line-of-sight, remains a major challenge in computer vision-based animal monitoring, particularly in enclosed environments where traditional tracking methods fail. This disrupts identity continuity, complicating long-term behavioural and health analyses. Building on prior work in 3D scene reconstruction via deep learning-based monocular depth estimation, we implement a relational quasi-3D bounding box re-identification method for cattle in a straw-bedded calving barn environment. By incorporating estimated depth, the method computes relative 3D positions and angles between detected animals, enabling spatial re-identification during misdetections and after occlusion. To improve temporal continuity, relational bounding boxes are used to interpolate missing detections and resolve ID inconsistencies. Initial tracking predictions were generated using a fine-tuned YOLOv8 model with ByteTrack for identity preservation. The re-identification (ReID) framework was evaluated on a 15-minute video (10,800 frames) and applied solely to identity reassignment and false detection recovery. Compared to the baseline tracker, the ReID-enhanced method improved identity consistency while maintaining high accuracy, achieving a Multi-Object Tracking Accuracy (MOTA) of 99.57% and IDF1 score of 99.79%, outperforming the baseline’s 99.48% MOTA and 99.78% IDF1. MOTA reflects overall tracking performance by accounting for missed detections, false positives, and identity switches, while IDF1 assesses identity consistency through the F1 score of correctly matched detections over time. Importantly, the identity switches dropped from 62 to 5, and false positives from 306 to 13, demonstrating stronger robustness under occlusion. This non-invasive, occlusion-resilient approach offers enhanced identification continuity over 2D-based tracking methods while relying on standard RGB camera systems.
| Period | 9 Jun 2025 |
|---|---|
| Event title | Artificial Intelligence 4 Animal Science: Advancing Digital Biomarkers with AI: Breakthroughs in animal identification, health and welfare monitoring, behavior analysis, and remote sensing technologies |
| Event type | Conference |
| Location | Zurich, SwitzerlandShow on map |
| Degree of Recognition | International |