STARS—Deep Diversified Analysis of Occlusion-Resiliency for Robot Vision in Medical Contexts

Occlusion is a critical challenge in Robot vision. Unfavorable circumstances including occlusions must be considered when integrating action recognition techniques into an autonomous robotic system. It becomes more critical when the recognition involves Medical scenarios. Occlusion of certain body parts can disrupt the geometric and temporal stability of recognition, significantly affecting accuracy. Despite its practical significance, occlusion in healthcare scenarios is rarely addressed by existing skeleton-based action recognition techniques. This study proposes STARS (Spatial Temporal Adaptive Recurrent Solution), a neural network-based technique designed for robust occlusion handling in medical robotics. The proposed method effectively handles temporal and spatial occlusions in medical scenarios, outperforming state-of-the-art techniques. The efficacy of the proposed model is evaluated using the NTU RGB+D dataset, processed through five schematic configurations: Fundamental, Temporal, Spatial, Calibration, and Appraisal. These configurations target occlusion in distinct ways. The findings reveal that occluded data samples impact recognition accuracy, and the proposed technique achieves results close to non-occluded scenarios. However, occlusion in medical activity detection requires further attention.