Recognition and interpretation of activities in videos

Abstract

Background subtraction is among the difficult activities in computer vision. Since thenature of the environment may involve some changes due to light effect, dynamic background,shadow, camouflage effect, etc. Many moving object detection methods are proposed,but the majority of them fails to handle the multi-modality of scenes. Based onthe Artificial Immune Recognition System, efficient background subtraction methods areproposed. In the first method, the Single Gaussian model is combined with the ArtificialImmune Recognition System to better represent pixel variations in the scenes thatcontain dynamic background. Artificial Immune Recognition System is used as a classificationtool that separates antigens represented by the foreground pixels from antibodiesthat modelled the background pixels. Each pixel in this proposition is modelled witha feature vector contains Gaussian attributes. As a second contribution, we have usedthe Artificial Immune Recognition System for managing the number of Gaussians dynamicallyin Gaussian Mixture Model instead to fix them a priori by the user. In thiscontribution, a set of new Gaussians is generated using two different strategies: the firstone (Random generation) uses the Artificial Immune Recognition System for improvingthe system decision, while in the second strategy (Directed generation), the ArtificialImmune Recognition System is used to improve the production of background models.To reduce the effect of brightness, each frame in the video sequence is transformed fromthe RGB to HSV color space. Artificial Immune Recognition System has also benefitedfrom some modifications to reduce research cost and to avoid the explosion of data in thememory cells set. For reducing the research cost on the most representative memory cellto the current antigen, the structure of the memory cells set is redefined as a binary tree(kd-tree). Furthermore, we have added two mechanisms to the basic Artificial ImmuneRecognition System to avoid the explosion of data in the memory cells set. The proposedmethods are implemented and tested on public datasets. The obtained results are largelysatisfactory compared to other state-of-the-art methods.