Real-time nonparametric background subtraction with tracking-based foreground update

Abstract A nonparametric real-time and high-quality moving object detection strategy in a GPU is proposed. To improve the quality of the results in sequences where the moving objects and the background have similar appearance, not only the background but also the foreground is modelled. Both models are constructed from spatio-temporal reference data to reduce false detections due to small displacements of the background, and to take into consideration the natural displacements of the foreground. To avoid using kernels with too large spatial widths, the spatial positions of the foreground reference data are updated at each new frame using a particle filter that is able to deal with an unknown and variable amount of regions. Additionally, an automatic selection of regions of interest is carried out, which allows reducing drastically the computational cost of both foreground and background models.  The proposed strategy has been validated using three databases containing many challenges for motion detection and the results have been compared to those of other state-of-the-art approaches.