Motion Based Bird Sensing Using Frame Differencing and Gaussian Mixture

Background segmentation is a general technique that aims at detecting the moving objects in a sequence of continuous scenes or video stream by separating the non-moving background from the moving foreground object. The success and weakness of a foreground and/or background segmentation method depends on several external factors such as the scene selection, lighting, and weather conditions. Hence, we perform a comprehensive quantitative and qualitative analysis of two such background segmentation methods frame differencing and Gaussian mixture for a continuously changing environment with articulated foreground objects. We try to detect birds as our foreground objects in image streams with varying backgrounds. This study can be used to evaluate the effectiveness of a segmentation technique in a constantly changing outdoor environment. We find that the Gaussian mixture approach is more accurate than the frame differencing approach; however, as a trade-off the Gaussian mixture approach takes far more time and memory to run. The segmentation results produced using these techniques are the foundation for any further analysis that biologists need to better understand bird motion, bird feeding habits, bird flight and other bird behaviors. Key Terms: 1.) Background segmentation – Segment the background from foreground. 2.) Frame differencing – Finding absolute difference between frames. 3.) Centroid – The center of a motion detected region. 4.) Ground truth – The actual data collected on site. F A C U L T Y m e n T o R Deborah Estrin Department of Computer Science, UCLA The Center for Embedded Networked Sensing (CENS) is a National Science Foundation Science and Technology Center. One of the driving applications at this center is in the automated sensing of ecosystem health indicators. Deep joined our lab in 2007 as a participant in our intensive summer undergraduate internship program. His principle task was assisting in the development of an image recognition program designed to identify avian activity. Specifically, Deep was able to detect birds utilizing techniques in frame differencing and image segmentation. The detection of birds in the natural environment is a critical step in aiding ecologist in the study of avian behavior through image sensors. Deep Shah was a delight to have in our program; his light-hearted humor combined with his determination to produce quality work made for a wonderful addition to our labs. Deep J. Shah1, Deborah Estrin2 Student Assistant: Afrouz Azari Graduate Student Assistants: Teresa Ko, Shaun Ahmadian, Mohammad Rahimi 1Department of Electrical Engineering, University of California, Riverside 2Department of Computer Science, UCLA