Attention-Based Occlusion Handling for Dynamic Human-Robot Interaction in Unstructured Environments

Abstract: Robust human detection and pose estimation under occlusion remain central unsolved challenges for robots operating in unstructured, cluttered, or dynamically evolving environments. Conventional perception pipelines degrade substantially when human body parts are partially or fully occluded by furniture, equipment, or other people, leading to unstable robot behaviour and potentially unsafe interactions. This paper proposes the Attention-Gated Occlusion- Robust Perception (AGORP) framework, a novel architecture that integrates spatial and channel attention with transformer-based contextual reasoning to enable real-time occlusion detection, body-part recovery, and intention prediction during human-robot interaction. The framework introduces a dual-stream encoder that processes RGB and depth modalities through cross-modal attention gates, coupled with an Occlusion-Aware Temporal Transformer (OATT) that exploits motion history to hallucinate occluded skeletal configurations. Experiments conducted on the JRDB, CrowdBot, and a newly collected indoor manipulation dataset demonstrate that AGORP achieves a mean Average Precision (mAP) of 74.3% on occluded pose estimation, a 12.7 percentage-point improvement over the strongest baseline, while sustaining 31 frames per second on an NVIDIA Jetson AGX Orin embedded platform. Qualitative and quantitative analyses confirm the framework’s generalisation across occlusion severities, scene densities, and lighting conditions.

Keywords: attention mechanism, occlusion handling, human-robot interaction, transformer networks, pose estimation, unstructured environments.