The detection of foreground objects in a video or image sequence is important for motion analysis or smart surveillance systems, as it allows them to distinguish between suspicious behavior and wandering humans or animals. Such research involves analyzing finite details of an image to separate foreground objects from the background; however, noise affects the ability to completely detect objects. The authors emphasize the dual modality of the images’ background, which aids in effectively separating the foreground objects, producing a complete and confined silhouette.

The dual-mode model detects the complete silhouette of a foreground object from the dynamic background by calculating grey levels of pixels occurring frequently in image sequences. According to the authors, this method is sensitive to both gradual and radical changes in the environment and is capable of reducing noise and flashes on the monitor created by the opening and closing of doors, for example. Further, the method extracts complete silhouettes of the foreground objects in a low-contrast image, without any noise removal or connected component analysis, and can adapt to changes in lighting and dynamic noise. This method considers monitoring both “indoor and outdoor scenarios with under-exposed environments and low-contrast foreground objects.”

For their experiments, the authors use Microsoft’s Wallflower dataset, which contains seven image sequence scenarios, including a moved object, time of day, light switch, waving trees, camouflage, bootstrapping, and foreground aperture. The authors compare the dual-mode model with single Gaussian, a mixture of Gaussians, kernel density estimation, and grayscale, arranging pair methods for performance measures. The authors observed larger memory usage with the proposed dual-mode model as opposed to the mixture of Gaussians.

The authors predict that “the dual-mode method can quickly respond to changes in illumination and accurately extract foreground objects against a low-contrast background.” The investigation of the dual modality of the background estimation and separation from foreground objects makes this paper worth reading. Future experiments will explore color videos for object detection.