Scholarlink

<p>Human pose estimation (HPE) is a vital perception capability for industrial vision in robot-inclusive manufacturing work cells, supporting safety monitoring, interaction analysis, and ergonomic assessment. Real-time HPE in assembly scenes remains challenging due to structured occlusions from benches, tools, and robot arms, irregular body orientations, and complex lighting, which frequently degrade upper-limb keypoints such as wrists and elbows. To improve the speed–accuracy trade-off under strict efficiency constraints, we propose YOLOv8-RADPose. This hybrid Transformer–CNN architecture integrates a lightweight Resolution-Aware Dense (RAD) Transformer block into the YOLOv8-Pose head to enhance global keypoint reasoning while preserving single-stage inference. RADPose introduces patch-free tokenization to retain fine-grained spatial correspondence, resolution-adaptive positional embeddings to support multi-scale processing without retraining positional parameters, and a triple-residual structure to stabilize optimization and preserve localization cues. On MS COCO, YOLOv8s-RADPose improves from 63.7 AP to 65.7 AP on test-dev2017 while running at 232.6 FPS on an NVIDIA RTX 4090 (FP16, batch size 1) with 16.9&#xa0;M parameters and 33.5 GFLOPs. The larger YOLOv8x-RADPose reaches 75.3 AP at 102 FPS, with 74.4&#xa0;M parameters and 267.4 GFLOPs, outperforming the corresponding YOLOv8x-Pose baseline&#xa0;by +3.2 AP. On CrowdPose, YOLOv8s-RADPose and YOLOv8x-RADPose achieve AP scores of 65.0 and 74.6, respectively, confirming consistent gains under crowded and occluded conditions. Quantitative and qualitative evaluation on the InHARD industrial dataset further indicates improved robustness to structured occlusions and challenging illumination in assembly-like scenes.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing industrial human pose estimation through a hybrid Transformer–CNN architecture

  • 期刊论文

  • 发表日期：

    2026年7月1日

查看全文
  • Muhammad Rashid,
  • Junfeng Wang,
  • Sulman Ahmed

摘要

Human pose estimation (HPE) is a vital perception capability for industrial vision in robot-inclusive manufacturing work cells, supporting safety monitoring, interaction analysis, and ergonomic assessment. Real-time HPE in assembly scenes remains challenging due to structured occlusions from benches, tools, and robot arms, irregular body orientations, and complex lighting, which frequently degrade upper-limb keypoints such as wrists and elbows. To improve the speed–accuracy trade-off under strict efficiency constraints, we propose YOLOv8-RADPose. This hybrid Transformer–CNN architecture integrates a lightweight Resolution-Aware Dense (RAD) Transformer block into the YOLOv8-Pose head to enhance global keypoint reasoning while preserving single-stage inference. RADPose introduces patch-free tokenization to retain fine-grained spatial correspondence, resolution-adaptive positional embeddings to support multi-scale processing without retraining positional parameters, and a triple-residual structure to stabilize optimization and preserve localization cues. On MS COCO, YOLOv8s-RADPose improves from 63.7 AP to 65.7 AP on test-dev2017 while running at 232.6 FPS on an NVIDIA RTX 4090 (FP16, batch size 1) with 16.9 M parameters and 33.5 GFLOPs. The larger YOLOv8x-RADPose reaches 75.3 AP at 102 FPS, with 74.4 M parameters and 267.4 GFLOPs, outperforming the corresponding YOLOv8x-Pose baseline by +3.2 AP. On CrowdPose, YOLOv8s-RADPose and YOLOv8x-RADPose achieve AP scores of 65.0 and 74.6, respectively, confirming consistent gains under crowded and occluded conditions. Quantitative and qualitative evaluation on the InHARD industrial dataset further indicates improved robustness to structured occlusions and challenging illumination in assembly-like scenes.