Improved Gaussian Mixtures for Robust Object Detection by Adaptive Multi-Background Generation Mahfuzul Haque, Manzur Murshed, and Manoranjan Paul Gippsland School of Information Technology, Monash University, Victoria 3842, Australia Email: {Mahfuzul.Haque, Manzur.Murshed, Manoranjan.Paul}@infotech.monash.edu.au Abstract Background Modelling Implemented System Frame 1 Frame 2 Frame t .. K i t i t i t t i t X w X P 1 , , , ) , , ( ) ( ) ( ) ( 2 1 2 / 1 2 / , , 1 | | ) 2 ( 1 ) ( t t T t t X X n t t t e X Gaussian Mixture Model (GMM) for each pixel Input scenes A pixel model is constructed and updated for each pixel which maintains a mixture of Gaussian distributions for modelling multi-modal distribution caused by moving foregrounds and repetitive background motions [1-3]. Adaptive Gaussian mixtures are widely used to model the dynamic background for real-time object detection. However, object quality still remains unacceptable due to poor Gaussian mixture quality, susceptibility to background/foreground data proportion, and instability with varying operating environments. This paper presents an effective technique to eliminate these drawbacks by modifying the new model induction logic and using intensity difference thresholding to detect objects from one or more believe-to-be backgrounds. The images shown in the header has been taken from http://www.informationliberation.com [1] Mahfuzul Haque, Manzur Murshed, and Manoranjan Paul, On Stable Dynamic Background Generation Technique using Gaussian Mixture Models for Robust Object Detection, IEEE International Conference On Advanced Video and Signal Based Surveillance (AVSS), New Mexico, USA, 2008. [2] Mahfuzul Haque, Manzur Murshed, and Manoranjan Paul, A Hybrid Object Detection Technique from Dynamic Background Using Gaussian Mixture Models, IEEE International Workshop on Multimedia Signal Processing (MMSP), Cairns, Australia, 2008. [3] Mahfuzul Haque, Manzur Murshed, and Manoranjan Paul, Improved Gaussian Mixtures for Robust Object Detection by Adaptive Multi-Background Generation, International Conference on Pattern Recognition (ICPR), Tampa, Florida, USA, 2008. Quantitative Evaluation First Frame Test Frame Ideal Result Lee’s Tech. Proposed Tech. P(x) Existing Models Intensity New Model New Model Induction Scheme 3 Experimental results on 14 test sequences including PETS and Wallflower datasets. Error rates at medium learning rate (α = 0.01) and the standard deviation of the error rates over three learning rates α = 0.1, α = 0.01, and α = 0.001. 1 2 8 6 Proposed Detection Scheme 4 Qualitative Evaluation 7 Model Quality Visualisation 5 One model Two models More than two models 0 127 255 Input Frames Visualisation Visual comparison results at medium learning rate, α = 0.01. Model matching: B/G Model Selection: F/G Detection: ; Model distance