Pose estimation is a computer vision technique that estimates the human body pose by localizing the main joints, such as elbows, knees, and ankles, known as the key points. There are plenty of pose estimation applications that are especially useful on mobile and embedded devices, including exercise pose correction.
In this work, we optimized a pose estimation model to deploy it on Jetson edge devices. The pose estimation model can work well on real-world CCTV data compared to the existing models.
How it works
We have used the open-source OpenPifPaf pose estimator model for this application. We exported the PyTorch OpenPifPaf model to ONNX format and used the exported model to generate a TensorRT inference engine deployable on Jetson platforms.
Since the OpenPifPaf model is designed for the transportation domain in crowded areas, it works well even on real-world CCTV videos where the human bodies only occupy a small portion of the video frame.
This is the first time that an OpenPifPaf model, which is a complex and heavy model, is deployed on Jetson devices. We faced many challenges along the way; we will share these challenges and the solution to each one with the community in our future blog posts.