In this study, one adaptive spatiotemporal background modelling algorithm is proposed for robust and reliable moving object detection in dynamic scene. First, a modified adaptive Gaussian mixture model (GMM) is presented to describe the temporal distribution of each pixel, based on which the spatial distribution of background is constructed by using non-parametric density estimation. By fusing the temporal and spatial distribution model, a heuristic strategy is presented for background subtraction. To reduce the computational cost, a novel criterion for adaptively determining the components number of GMM and the integral image method for calculating the spatial distribution model are proposed. Several experiments show that the proposed method can effectively reduce false positives caused by sudden or gradual changes of the background, and maintains lower false negatives, compared with some representative algorithms.
ASJC Scopus subject areas
- Computer Vision and Pattern Recognition