Segmentation of Palmprint Regions from Diversified Hand Videos

Palmprint biometrics has been a robust medium of identifying human subjects. The segmentation of palmprint regions however remains less explored, especially when it comes to real-word acquisition conditions. This paper presents an efficient palmprint segmentation framework using a proprietary hand video dataset captured with diverse backgrounds and other representational variations including distance from sensor, hand orientation, change in illumination, deformation and finger adjustments. This indigenous video hand dataset comprises of videos from left and right hands of 30individuals and containing common occlusions like rings and other accessories. The proposed framework works by developing a viable ground truth through polygon type annotations. The Segment Anything Model is employed with prompts of points from the developed ground truth to segment the palmprint region from the video frames, which can further be utilized to attain discerning palmprint features leading to robust biometrics system.