also different th<strong>an</strong> using <strong>an</strong> un<strong>an</strong>imous vote scheme for classifier fusion in the ensemble classifierpresented in this work. This is because if only a few features match the new inst<strong>an</strong>ce the classifierswill update themselves based on a match in the ensemble algorithm, while for the single featurevector algorithm the update would not be based on a match unless every dimension was matched(where matching is being <strong>with</strong> 2.5σ).Evaluation <strong>of</strong> the ensemble algorithm’s perform<strong>an</strong>ce could be performed using the VACE evaluationframework in [57]. This framework <strong>of</strong>fers a more comprehensive <strong>an</strong>alysis <strong>of</strong> the algorithmsrobustness, as well as a comparison against the perform<strong>an</strong>ce <strong>of</strong> other tracking algorithms.Real-time implementation <strong>of</strong> this system would require a system <strong>with</strong> multiple processingcores <strong>an</strong>d a large sized main memory, but otherwise its realization is believed to be reasonable.Enh<strong>an</strong>ced engineering <strong>of</strong> the algorithms to exploit the inherent parallelism <strong>of</strong> the ensemble shouldallow for the h<strong>an</strong>dling <strong>of</strong> relatively high frame rates depending on the image resolution.6.3 Final ThoughtsIt is believed that further research needs to be conducted in order to further underst<strong>an</strong>d thegains that were observed in this work. There was a clearly demonstrated adv<strong>an</strong>tage to usingadditional features in background classification, even though m<strong>an</strong>y <strong>of</strong> these features are verypoor when used individually. Through greater investigation a more clearly defined set <strong>of</strong> optimalensemble features would allow for a more general background classification algorithm that couldhave key impacts in defense, robotics, <strong>an</strong>d surveill<strong>an</strong>ce applications.56
REFERENCES[1] R. Jain, R. Kasturi, <strong>an</strong>d B. Schunck. Machine Vision. McGraw Hill, 1995.[2] C. Wren, A. Azarbayej<strong>an</strong>i, T. Darrell, <strong>an</strong>d A. Pentl<strong>an</strong>d. Pfinder: Real-time tracking <strong>of</strong> thehum<strong>an</strong> body. IEEE Tr<strong>an</strong>s. Pattern Anal. Mach. Intell., 19(7):780–785, 1997.[3] D. Koller, J. Weber, <strong>an</strong>d J. Malik. Robust multiple car tracking <strong>with</strong> occlusion reasoning.In ECCV (1), pages 189–196, 1994.[4] C. Stauffer <strong>an</strong>d W. Grimson. Adaptive background mixture models for real-time tracking.In IEEE Conference on Computer Vision <strong>an</strong>d Pattern Recognition, volume 2, 1999.[5] J. B. MacQueen. Some methods for classification <strong>an</strong>d <strong>an</strong>alysis <strong>of</strong> multivariate observations.In Proceedings <strong>of</strong> 5th Berkeley Symposium on Mathematical Statistics <strong>an</strong>d Probability, pages281–297, 1967.[6] S. Atev, O. Masoud, <strong>an</strong>d N. Pap<strong>an</strong>ikolopoulos. Practical mixtures <strong>of</strong> gaussi<strong>an</strong>s <strong>with</strong> brightnessmonitoring. In The 7th International IEEE Conference on Intelligent Tr<strong>an</strong>sportationSystems, 2004.[7] G. Gordon, T. Darrell, M. Harville, <strong>an</strong>d J. Woodfill. <strong>Background</strong> estimation <strong>an</strong>d removalbased on r<strong>an</strong>ge <strong>an</strong>d color. In Image Processing, 2001. Proceedings. 2001 International Conferenceon, volume 2, pages 395–398, 2001.[8] M. Harville, G. Gordon, <strong>an</strong>d J. Woodfill. Adaptive video background modeling using color<strong>an</strong>d depth. In Conference on Image Processing, pages 90–93, 2001.[9] K. Karm<strong>an</strong>n, A. Br<strong>an</strong>dt, <strong>an</strong>d R. Gerl. Moving object segmentation based on adaptivereference images. In Europe<strong>an</strong> Signal Processing Conf, pages 951–954, 1990.[10] C. Ridder, O. Munkelt, <strong>an</strong>d H. Kirchner. Adaptive bbackground estimation <strong>an</strong>d foregrounddetection using kalm<strong>an</strong> filtering. In In Proc. ICAM, pages 193–199, 1995.[11] R. E. Kalm<strong>an</strong>. A new approach to linear filtering <strong>an</strong>d prediction problems. Journal <strong>of</strong> BasicEngineering, 82:3545, 1960.[12] M. Leung <strong>an</strong>d Y. Y<strong>an</strong>g. Hum<strong>an</strong> body motion segmentation in a complex scene. PatternRecognition, 20(1):55–64, 1987.[13] K. Toyama, J. Krumm, B. Brumitt, <strong>an</strong>d B. Meyers. Wallflower: Principles <strong>an</strong>d practice <strong>of</strong>background mainten<strong>an</strong>ce. In ICCV (1), pages 255–261, 1999.[14] Y. Zh<strong>an</strong>g, Z. Li<strong>an</strong>g, Z. Hou, H. W<strong>an</strong>g, <strong>an</strong>d M. T<strong>an</strong>. An adaptive mixture gaussi<strong>an</strong> backgroundmodel <strong>with</strong> online background reconstruction <strong>an</strong>d adjustable foreground mergence time formotion segmentation. In IEEE International Conference on Industrial Technology, volume14-17, pages 23–27, 2005.57