Video surveillance has become ubiquitous in our lives, causing both beneficial and harmful results. However, the trend of increasing video surveillance will continue as long as the technology for automated processing of video streams improves. Background subtraction, one of the most popular components of the automated content analysis of video data, removes the stable background from the data so that moving foregrounds can be extracted and analyzed. The problem is that the background is often not as stable as it seems, due to noise, illumination changes, and the alteration of background objects. This paper proposes a new method to cope with illumination changes to the background.
The technique needs images of the background illuminated with different lighting conditions. These images are clustered, and then a probabilistic model of the background is computed for each cluster using principal component analysis. The model forms a subspace in a larger vector space of images. This calibration step has to be done before the system is deployed. During the surveillance mode, the closest subspace for each incoming video frame is identified. The projection to the subspace is considered the background. By subtracting the background from the video frame, the system extracts the foreground.
As the authors acknowledge, there are a few issues with the method, including the handling of background alteration and determining the number of clusters. The paper fails to mention that the calibration step may not be possible in uncontrolled environments such as cities and highways. A more incremental approach to the calibration is required.
The authors’ approach should be more effective in handling illumination changes than a standard approach that uses a single subspace. Overall, the paper is well organized and easy to follow.