A Low False Negative Filter for Detecting Rare Bird Species from Short Video Segments using a Probable Observation Data Set-based EKF Method

Dezhen Song, Yiliang XuAAAI 2010 · DOI

Introduction

Ornithologists searching for rare or thought-to-be-extinct species, such as the Ivory-billed Woodpecker (IBWO), face a massive data challenge. Automated camera systems in remote habitats like Bayou DeView, Arkansas, generate terabytes of video footage. However, a target bird may only appear for a fraction of a second, a few times a year. Standard motion detection is often too sensitive, triggering on falling leaves or insects, while traditional tracking algorithms like the Extended Kalman Filter (EKF) often fail because the observation windows are too short and measurement noise is too high. In this context, a false negative—missing a real sighting—is a catastrophic failure.

Methodology: The PODS-EKF Approach

To address this, researchers developed the Probable Observation Data Set (PODS)-based EKF. This method models a bird as a 3D line segment flying in a straight line at a constant velocity, mapped to 2D image coordinates via a pin-hole camera model.

Figure 2: Geometric model mapping 3D flight trajectories to the 2D image plane.

Unlike a standard EKF that attempts to estimate a state from noisy measurements, the PODS approach defines an error range around every detected point. The algorithm searches this entire error space to find any set of points (the Probable Observation Data Set) that allows a standard EKF to converge on a physically plausible flight trajectory. By searching for the "best-case" interpretation of noisy data, the filter ensures that if a target bird actually flew by, it will be flagged for human review.

Experimental Results and Data Reduction

The system was validated using a combination of 1,000,000 simulated trajectories, controlled field tests with Rock Pigeons, and a massive 29.41 TB archive of raw field video.

Figure 3: (a) Convergence comparison between standard EKF and PODS-EKF; (b) False positive (FP) and false negative (FN) rates across different species.

Key performance metrics include:

  • Zero False Negatives: In field tests, the algorithm caught 100% of confirmed bird flights.
  • Massive Data Compression: When applied to the year-long IBWO search archive, the filter reduced nearly 30 TB of raw data to just 146.7 MB—a reduction rate of 99.9995%.
  • High Accuracy: The system achieved a 95.0% Area Under the ROC Curve (AUC) in physical experiments, significantly outperforming standard EKF implementations which often diverge under high noise.

Discussion and Limitations

The PODS-EKF acts as a high-pass filter, effectively removing the "human-in-the-loop" bottleneck by presenting experts with only the most plausible candidates. While highly effective, the current model assumes straight-line flight at constant speeds, which may not account for maneuvering birds. Additionally, the monocular setup faces challenges when birds fly directly toward or away from the camera, as the resulting image size changes are less distinct than lateral motion. Future work may involve modern hardware acceleration to handle higher resolution and frame rates in real-time.

Summary updated

BibTeX
@inproceedings{song2010low,
  author = {Dezhen Song and Yiliang Xu},
  title = {{A Low False Negative Filter for Detecting Rare Bird Species from Short Video Segments using a Probable Observation Data Set-based EKF Method}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2010},
  doi = {10.1609/aaai.v24i1.7722},
}
Figure