Bird Species Detection using Monocular Vision System

Introduction

Our group focuses on developing autonomous observatories to assist nature scientists to search rare birds in remote environments such as in a deep forest. Due to power and communication constraints, it is often prohibitive to install dislocated multiple cameras in deep forest to form stereo pairs. Also, the calibration and the precise synchronization of dislocated stereo cameras are very difficult. Therefore, monocular vision is preferred in such settings. Since the final bird identification has to be performed by human experts, it is necessary to develop a filter to reduce the huge volume of the video data to a manageable size for human experts. Since the chance of the rare bird flying in front of the camera is very small and missing the opportunity would be costly, the system requires an asymmetric filter design that emphasizes very low false negative with manageable false positive.

Fig. 1. Camera installed in deep forest

 

Fig. 2. Bird motion sequence

The input of the problem is a segmented motion sequence of an object from consecutive video frames. The output of the problem is to determine whether the motion sequence is caused by a targeted bird species.

 

Bird Modelling

To address the problem, we study the bird flying data and find that the motion of a flying bird incorprates both translation motion and periodic motion. A bird body axis is an invariant dimension during flying and the bird body axis is often parallel to that of the tangent line of the bird flying trajectory. We also find bird wing-flapping has uniqe narrow frequence range during steady flight for each bird species. We model the flying bird to capture both body axis translation motion and the wing-flapping periodic motion. A bird filter algorithm is built based on the modelling to extract flying bird's translation and periodic motion features (e.g., speed, wingbeat frequency) and compare them with the prior known profile of the targeted bird.

Fig. 3. Modeling of bird translation motion

 

 

 

Fig. 4. Modeling of bird periodic motion

 

Experiments

a) Detection based on translation motion

We have implemented and tested the bird filter algorithm based on bird body axis translation motion by both the simulated data and the real data from field experiments. We chose Arecont Vision 3100 high resolution networked video cameras as the imaging devices. The camera runs at 11 frames per second with a resolution of 3 Mega-pixel per frame. The lens for the camera is a Tamron auto iris vari-focus lens with a focal length range of 10-40 mm. The algorithm achieves very low false negative rate, which is crucial for our bird search purpose. The false postive rate is manageable. The bird detection algorithm has helped us reduce the video data by 99.9994%, which were captured during Oct. 2006 to Oct. 2007 in a deep forest in eastern Arkansas to search for the thought-to-be-extinct ivory-billed woodpeck (IBWO). The algorithm also achieves 95.3% area under the ROC curve in the physical experiment for detecting rock pigeons.

 

Fig. 5. False positive and false negative rates for house sparrow, rock pigeon and ivory-billed woodpecker (IBWO) with differnt thresholds

 

 

Fig. 6. ROC curves for rock pigeon in both simulation and physical experiment

 

 

b) Detection based on periodic motion (coming soon)

 

Publications
  1. Dezhen Song and Yiliang Yu, A Low False Negative Filter for Detecting Rare Bird Species from Short Video Segments using a Probable Observation Data Set-based EKF Method, IEEE Transactions on Image Processing (Accepted with minor revision) (PDF 1300K).
  2. Dezhen Song, Ni Qin, Yiliang Xu, Chang Young Kim, David Luneau, and Ken Goldberg, System and Algorithms for an Autonomous Observatory Assisting the Search for the Ivory-Billed Woodpecker, IEEE International Conference on Automation Science and Engineering (CASE), Washington DC, August, 2008 (PDF 1400K).