AAmogh·Panhale
All projects
Computer VisionML2026

Wildlife CV Tracking

Classical computer-vision trackers benchmarked for monitoring wildlife in video.

Code

Problem

Wildlife monitoring needs reliable object tracking across video, but animals moving in groups and changing appearance break naive trackers. Built for CMPUT 428 at the University of Alberta.

What I built

  • Implemented and compared four classical trackers: Inverse Compositional Lucas-Kanade, mean-shift, MOSSE, and a discrete Kalman filter.
  • Evaluated each on the AnimalTrack and Zebra datasets by IoU and FPS against ground-truth annotations.
  • Shipped a Streamlit demo to explore tracker outputs and performance comparisons interactively.

Impact

  • MOSSE led on IoU and FPS, with IC Lucas-Kanade close behind; found mean-shift degrades on grouped animals.
  • Identified the next step — a hybrid detector + re-identification (YOLO) strategy — from the failure analysis.

Stack

PythonOpenCVNumPyStreamlit