I would like to work on improving background subtraction algorithm. In particular, I want to implement algorithm from the following article: Background Subtraction using Local SVD Binary Pattern, Guo et al. (2016)
It is based on LSBP feature descriptors and achieves state-of-the-art performance on the CDnet 2012 dataset. LSBP descriptors are particularly good in regions with illumination variation, noise and shadows. So, this algorithm has better performance in this kind of regions.
In addition, this algorithm has another pleasant trait: after extraction of LSBP descriptors it processes frames pixel-wise (i.e. independently). This is the common trait among all the consensus-based background models. It means, that implementation can leverage the parallelism inside OpenCV. Furthermore, implementation could be fully parallelized and implemented on both CPU/GPU. Thus, this algorithm will be fast enough for the realtime processing even on mobile devices.