Reliable background prediction using approximated GMM
Tomosuke Maeda, Tomohiko Ohtsuka · 2015
Our study proposes a new reliable background prediction for object detection in a frame sequence. Our method generates the approximated Gaussian Mixture Model (GMM) from the standard GMM by eliminating moving objects that can be easily detected based on frame differences. This reduces the computational time taken to predict the background image by averaging the intensity of each pixel of approximated GMM. However, the computational time costs more to fit each GMM parameter using an EM algorithm. In addition, this method achieves a reliable background prediction. This is possible because the precision of the background prediction is higher than other conventional approaches. Using the proposed background subtraction method, our experimental results indicate that the precision and recall levels obtained were approximately 20% higher than other levels that were obtained using conventional approaches.