Computer Vision Team
FarmGuard
Tracking Architecture
We combine YOLOv5 detections with Kalman and Extended Kalman Filters in a dual-path pipeline for robust deer tracking.
How the Pipeline Works
-
Input video
Raw drone footage feeds into our vision system. -
YOLO v5
Rapid object detector pinpoints deer bounding boxes. -
Deer Detection
Aggregates YOLO proposals, filters out false positives. -
Tracked Deer
Outputs stable positions for the downstream module. -
Kalman Filter
Smooths the raw position sequence, assuming constant velocity. -
Extended Kalman Filter
Handles non-linear motion and camera rotations for extra robustness.
YOLOv5 Detection
Detects deer bounding boxes each frame, providing the raw measurements for tracking.
Kalman Filter
Assumes constant velocity to predict and smooth positions over time, reducing jitter.
Extended Kalman Filter
Incorporates camera rotation and non-linear motion; matches detections via Mahalanobis distance.
Live Demos
