A DenseNet-OpenPose Framework for Static Biometric Recognition via Human Body Keypoints

The conventional biometric modalities, including fingerprints, iris patterns, and facial features, have been considered for recognition during the past few years. To increase the recognition accuracy, we propose a novel biometric recognition framework to consider more static anthropometric parameters, which include body shape, height, arm span, and limb proportions. In addition, we propose a DenseNet-enhanced OpenPose architecture that is developed to extract 2D skeletal keypoints from RGB images, incorporating a Coordinate Attention (CA) mechanism to enhance keypoint localisation accuracy. Based on the extracted keypoints, robust static measurements and morphological descriptors are derived to construct discriminative biometric features. The proposed model is evaluated on the Southampton Gait Database using both Object Keypoint Similarity (OKS)-based precision and classification accuracy metrics. Experimental results demonstrate that even under conditions involving varied viewpoints and partial occlusions. The proposed approach offers a privacy-preserving alternative for biometric recognition, with promising applications in scenarios such as healthcare monitoring, non-cooperative surveillance, and contactless identity verification.