In this proposal, a new approach for background subtraction by using multiple saliency methods and saliency optimization frameworks is proposed. Two proposed bottom-up saliency methods use low level cues such as boundary connectivity, local contrast of the image areas to their surroundings and sliding window approach. The other proposed top-down saliency method uses high level cues pretrained on ImageNet. Three proposed saliency methods are all better than current OpenCV saliency module. With the result from multiple saliency methods, background subtraction can be performed by using exponentially decaying method or aggregated saliency optimization framework. The proposed approach tend to be more robust when the background is highly dynamic or the camera is in motion compared with current OpenCV Background Subtraction module.

Organization

Student

SHENGXIN QIAN

Mentors

  • Antonella Cascitelli
close

2017