Human Keypoint Recognition

Human keypoint recognition is an AI algorithm that detects 17 joint points (keypoints) on the human body from images or video. It is utilized across a wide range of applications requiring pose estimation, including sports motion analysis, fall detection in nursing care settings, and work posture assessment in factories.

Algorithm Overview

Using deep learning-based pose estimation models, 17 keypoints on the human body are detected. Each keypoint corresponds to the following body parts:

Index	Body Part	Index	Body Part
0	Nose	9	Left Wrist
1	Left Eye	10	Right Hip
2	Right Eye	11	Left Hip
3	Left Ear	12	Right Knee
4	Right Ear	13	Left Knee
5	Right Shoulder	14	Right Ankle
6	Left Shoulder	15	Left Ankle
7	Right Elbow	16	—
8	Left Elbow	—	—

Performance Metrics

Model	Accuracy (mAP pose@0.5)
Person Pose-S (Lightweight)	86.3
Person Pose-M (Standard)	89.3

Edge AI Board (RV1126B) Execution Efficiency

Model	Processing Time
Person Pose-S (Lightweight)	59ms
Person Pose-M (Standard)	103ms

Key Features

17-point joint detection: High-precision detection of major body joints
Two-tier model selection: Choice between speed-priority (S) and accuracy-priority (M)
Real-time processing: Edge inference at approximately 59ms with the lightweight model
Foundation for pose estimation: Pre-processing for fall detection, motion analysis, and sports form analysis

Use Cases

Fall detection in nursing care facilities (identification of unnatural postures)
Sports training form analysis
Factory work posture assessment (back pain prevention)
Rehabilitation motion recording
Gesture interfaces
Retail store customer behavior analysis

Edge AI Board Implementation

Using the RV1126B NPU, keypoint detection is achieved with 59ms (lightweight model) and 103ms (standard model). Camera video is processed in real time, enabling immediate detection of postural anomalies.