Unsupervised Moving Target Detection in Dynamic Scenes

摘要

We present an unsupervised algorithm for detection of moving targets in highly dynamic scenes. These are scenes whose background is subject to stochastic motion, due to the presence of multiple moving objects (crowds), water, trees swaying in the wind, etc. The algorithm is inspired by biological vision. Target detection is posed as a problem of center-surround saliency, which aims to identify the locations of the visual field of maximal contrast with the background. Contrast is defined in terms of both appearance and motion dynamics, and measured using mutual information between stochastic models, known as dynamic textures, which can account for complex motion. This enables very robust target detection in the classes of scenes which have traditionally proven most adverse to tracking. Extensive tests in the context of dynamic background subtraction have shown significantly superior performance to previous techniques.