Implement Skeleton Detection Model (RTMO + RealSense Integration)

[Hongyi-Sam-Dong]

Skeleton Detection — RTMO + RealSense Integration

Implemented a real-time multi-person skeleton detection pipeline on the DGX Spark using RTMO-M and Intel RealSense D456.

[Hongyi-Sam-Dong]

Current functionality:

• RealSense D456 captures RGB at 848×480 @ 60 FPS
• RGB frames are accessed directly through pyrealsense2
• RealSense capture and RTMO inference run in the same Python process
• Large RGB frames do not pass through ROS/Fast DDS before inference, avoiding the previous image transport bottleneck
• RTMO outputs:
• person bounding box
• person confidence
• COCO-17 body keypoints
• skeleton connections
• Results are published as custom ROS 2 SkeletonFrame messages
• Uses a latest-frame-wins buffer to avoid accumulating stale frames
• Current measured performance:
• Camera: ~60 FPS
• Skeleton output: ~50 FPS
• RTMO inference: ~20 ms/frame
• Offline/debug skeleton visualization is also supported

[Hongyi-Sam-Dong]

Current Pipeline:
Intel RealSense D456
│
│ pyrealsense2
▼
RGB Frame (BGR NumPy)
│
│ same process
▼
RTMO-M
│
▼
Person Detection
( COCO-17 Keypoints
Bounding Box
Confidence)
│
▼
SkeletonFrame ROS Message
│
▼
/skeleton_detection/frame