UC Riverside

Latest research directions are trying to explore recognition systems that can handle challenging covariates like change in clothing, pose variations, distortions due to range, and occlusion. Given the challenges in recognizing humans in these adverse conditions, 3D models seem to be a promising approach in dealing with these challenges in a robust manner. This is because 3D models try to capture a person’s unique physical attributes like their body shape and joint distances/ratios which are unique to every individual and remain relatively unaffected by variations in clothing, lighting conditions, pose, range/distortion, and occlusions. The challenge lies in accurately extracting a 3D model from a single image or a video such that the 3D model accurately encapsulates the identity preserving characteristics like shape, pose and joint information of the person. Due to advancements in deep learning techniques and hardware capabilities, exciting research has lately been produced on generating 3D models from a single image or a video, but all this work focuses only on capturing poses and general body shape of the human being for applications/use in virtual or augmented reality. Limited work exists on making the 3D model identity preserving so that it can be used for human recognition. Existing 3D model-based recognition systems make use of a subject’s gait cycle which requires them to use multiple frames from a single video and allow them to leverage temporal information. In this thesis, we explore a single RGB image-based 3D model generation of an individual that can be used for identification of the person using body-based characteristics like the shape, pose and joint information. We investigate the methods to make 3D models identity preserving and develop a deep learning-based approach for human recognition under the challenging covariates mentioned above.