Motion Detection Based On Spatiotemporal Saliency In Dynamic Scenes

Salience detection is one of the research focus in computer vision, which simulates human visual attention and vision information processing mechanisms to found saliency computation models. So far, for the static image, there have been many mature methods for saliency regional extraction. Saliency detection focuses on local feature contrast rather than global feature changes. So it can avoid the problem of motion compensation in dynamic scenes.Motion detection is a classical problem in computer vision research. Highly dynamic background and camera motion are two challenges of this task. Spatiotemporal saliency detection has been used in motion detection and is robust to the challenges.This paper reviews the saliency research achievement, and generally divided these algorithms into three categories: algorithms based on low-level features, algorithms based on image complexity and algorithms based on biological visual model. Then introduces some important biological basis related with saliency, based on which we propose 4 basic principles of saliency algorithm design. This article also reviews the classic motion detection algorithms, and presents a spatiotemporal saliency based on mixtures of textures which then is used to detect moving object in dynamic scenes.The main work and innovations are:(1) reviewed the significant progress in algorithm research at home and abroad, the advantages and disadvantages of commonly used algorithms and scenarios;(2)Implement some classic moving target detection algorithms, such as temporal difference, W4, and GMM, and make some improvements by analyzing our experiments;(3) Based on the mechanisms of human visual system, the paper summarizes 4 principles for the saliency algorithms design;(4)We present a novel spatiotemporal saliency detection method based on mixtures of textures(MDT), which is then used to resolve the motion detection problem. First we model the highly dynamic background by mixtures of textures. Then calculate the spatiotemporal saliency based on a center-surround framework. Finally the saliency map yields motion detection results after thresholding. The experiments demonstrate that our method can effectively promote the accuracy of motion detection.