(I) Objective: Develop a pose estimation system to identify key points of a child sitting at a desk, enabling the detection of unhealthy sitting postures such as:

• Lowering their head or lying on the table. (No. 2 in the right image)

• Tilting their head. (No. 3 in the right image)

• Leaning their head on their left/right arm. (No. 4,5 in the right image)

• Supporting their head with their left/right hand. (No. 6,7 in the right image)

(II) Approach: Utilized the YOLOv7-Tiny Pose model to meet real-time frame-per-second (FPS) requirements on low-compute-power chips in the company’s devices.

(III) Dataset: Collected images of children sitting at desks from client-provided data, ensuring diverse representations of healthy and unhealthy sitting postures.

(VI) Results:

• The YOLOv7-Tiny Pose model effectively recognized a wide range of healthy and unhealthy sitting postures.

• Demonstrated reliable and efficient performance, supporting real-time monitoring in practical applications.

The below images show the YOLOv7-Pose model's predictions, including person bounding boxes (blue), head bounding boxes (red), hand bounding boxes (green), and key points.