A MULTISCALE SPATIO-TEMPORAL BACKGROUND MODEL FOR MOTION DETECTION

In this paper, we present a multiscale background model for motion detection. The proposed approach follows a nonparametric background modeling paradigm: each location in a dynamic scene collects a set of samples on different spatial scales which occurred in the past time and in the neighborhood. The motion measure of a location on a certain spatial scale hinges on how many samples existed in its context set are perceivably different from the sample at the same location of the incoming frame. The propagation of motion measure across scales and the soft updating scheme make this model applicable to dynamic background. We evaluate the proposed multiscale background model on a benchmark dataset which consists of nearly 90,000 frames in 31 videos representing 6 categories, and the experimental results demonstrate that it can efficiently detect motion in low contrast dynamic scenes.