trt_pose is an NVIDIA project for real-time human pose estimation on Jetson devices, accelerated with TensorRT. It detects body keypoints — eyes, elbows, ankles, and the rest of the standard keypoint set — fast enough to run live on a Jetson Nano. The repository is MIT licensed, has earned over 1,000 stars on GitHub, and ships with pre-trained models so you can try it immediately.
Pose estimation is usually too heavy for small edge boards; trt_pose attacks that problem by optimizing PyTorch models through torch2trt and TensorRT, so inference stays real-time on Jetson Nano and Jetson Xavier hardware. Beyond the demos, the project includes training scripts that work with any keypoint dataset in MSCOCO format, which means you can retrain it for hands, animals, machinery, or other keypoint tasks beyond the standard human skeleton. A companion project, trt_pose_hand, extends the same approach to hand pose and gesture recognition. Coming from NVIDIA’s AI-IoT team, the code is written as a practical starting point for embedded vision products rather than a research prototype.
Practically, trt_pose lowers the barrier to entry for edge pose estimation: the pre-trained models and examples let you go from a fresh Jetson board to a live skeleton overlay on camera input quickly, without first collecting data or training anything. From there, the included training pipeline and the torch2trt conversion path give you a route to customize — swap in a model trained on your own keypoints, convert it for TensorRT, and deploy it back on the same inexpensive hardware for interactive installations, robotics perception, or sports and rehabilitation prototypes.
trt_pose is free and open source under the MIT license, including its pre-trained models; you supply the Jetson hardware it runs on.
It is aimed at robotics, kiosk, fitness-tech, and interactive-installation builders who need camera-based body tracking on low-power edge devices rather than in the cloud.