Automated human recognition is a difficult challenge in using incomplete faces in bio-metric computer vision. As a result, periocular detection aims to discover humans by utilizing characteristics derived from the area around the eye. The region bounded by the half of the nasal region, jawline, and apex of the brow is used for periocular identification. As seen, the periocular facial structure comprises eye edges, eyebrows, eye foldings, and texture. Addressing variability in dynamic periocular identification is required due to differences in light, topic distances, sensor variances, and indoor-outdoor variations. To address this research difficulty, a HOG-based gradient approach for training deep CNN models is presented, which aids in the creation of domain invariant embedding space.