Background estimation is an important task in computer vision which involves using motion to isolate object(s) of interest in videos. A recent algorithm called SuperBE uses superpixels (groups of pixels clustered together for colour and spatial coherency) to help improve accuracy and reduce computational complexity.
We are interested in exploring ways to further optimise the algorithm, for speed and accuracy, and to compare this against other state-of-the-art approaches.
Our goals in this project are to attempt further simplification of the existing algorithm without significant accuracy loss and port the algorithm to a true-parallel embedded platform containing ARM-FPGA SoCs. We would then like to improve the accuracy of the algorithm on the FPGA system through researching additional and/or alternative discriminative features to be used in our algorithm’s background estimation process as well as general algorithm improvement for parallel performance. We would also like to consider running the algorithm on a GPU based platform to explore the options of speeding up the process of background estimation and possibly comparing it to the FPGA solution.
Undergraduate
None
Lab allocations have not been finalised