Monocular Vision-based Detection of a Flying Bird

To assist nature observation, we take on the challenge of detecting the species of a flying bird using a single camera. We study the bird flying data and find that a bird body axis is an invariant dimension during flight. We then develop a model-based detection approach that verifies the body axis information with the known bird flying dynamics. As a commonly used method, an extended Kalman filter (EKF) cannot be directly applied because the EKF would not converge due to the high measurement error introduced by image segmentation and the limited observation data due to the high flying speed of the bird. To cope with the problem, we develop a novel Probable Observation Data Set (PODS)-based EKF method. First, we prove that the EKF converges when there is no measurement error, and the new PODS-EKF searches the measurement error range for all probable observation data that ensures the convergence of the corresponding EKF. The detection is based on whether the set PODS is non-empty and the corresponding velocity is within the known bird flying velocity profile. The algorithm has been extensively tested using both simulated inputs and physical experiments. The results are satisfying and have shown the bird detector has less than 7% false negative rate and 90% area under the receiver operating characteristic (ROC) curve.