Moving Objects Detection Based On An Adaptive Background Subtraction Method

In the computer vision field, background subtraction methods are widely usedin segmenting the foreground targets in video sequences. However, with the rapiddevelopment of the computer vision applications, early unimodal backgroundsubtraction methods which just rely on the temporal domain to segment objectscannot satisfy the growing complex changes of the scene. The self-organizingbackground subtraction adopts several neurons to model one of the backgroundpixels. In addition, the method combines spatial information with temporalinformation by means of propagating context recursively. This method achievesgood detection performance in many video scenes. On the other hand, thedetermining of the parameters is very cumbersome and the adaptability of thediscrete changes in the background is inadequate. This paper proposes an adaptivemethod that improves the performance of the self-organizing backgroundsubtraction.As the self organizing background subtraction method depends on the initialparameterization too much, the research proposes an adaptive method whichcalculates the learning rate dynamically according the feedbacks of the detection.The proposed method avoids the selection of calibration frames artificially.Therefore, it simplifies the setting of the parameters and improves the convergencerate of the neurons.The research adopts the pixel level verification, which is based on the variationdetection of the local edge and the object level verification, which is based on theforeground contours matching to solve the deadlock and the ghostsphenomenon. The pixel level verification method can maintain the continuity of thebackground model in the spatial so that it has a better stability. On the other hand,the object level verification method is effective in eliminating the ghosts. Whereas,the completely discrete updating used in the method sacrifices the continuity of thebackground model and results in the loss of the stability. The results show thatobject level verification achieves greater performance when the video sequencescontain discrete background changing. However, pixel level verification does betterin the scenes that change steady. Both the two method improve the adaptive abilityto the changes of the background for the self-organizing background subtractionmethod.