People Detection
People detection is an AI recognition algorithm that identifies the position of people in images or video. It is an essential pre-processing technology for various person-related AI tasks such as person recognition, person tracking, and behavior analysis.
Algorithm Overview
People detection uses deep learning-based object detection methods to detect the position and extent (bounding box) of people in images with high precision.
It provides robust detection performance that is resistant to posture changes, clothing diversity, lighting conditions, and occlusion. Compared to general object detection, people detection is characterized by smaller aspect ratio variations but larger scale variations (from a few pixels to the entire image).
Performance Metrics
| Dataset | Detection Accuracy |
|---|---|
| FDDB | 98.64% |
Edge AI Board (RV1126B) Execution Efficiency
| Algorithm | Processing Time |
|---|---|
| face_detect | 28ms |
Key Features
- High-precision detection: 98.64% detection accuracy on the FDDB benchmark
- Diverse environment support: Stable operation indoors/outdoors, day or night
- Real-time performance: 28ms fast inference on edge AI boards
- Simultaneous multi-person detection: Capable of detecting multiple people in a single frame
Use Cases
- Entry/exit management systems (pre-processing for face recognition)
- Surveillance camera person detection
- Occupancy monitoring in nursing care facilities
- Store visitor counting
- Person tracking and movement flow analysis
- Pre-processing for facial attribute classification and expression analysis
Edge AI Board Implementation
Leveraging the RV1126B NPU (2.0 TOPS), low-latency inference of 28ms is achieved. In combination with a camera module, standalone edge AI people detection is possible.