Human activity recognition at a kilometer range using single-photon LiDAR.

Area of Science:

• Computer Vision
• Robotics
• Sensor Technology

Background:

• Human activity recognition is crucial but faces challenges with traditional RGB methods.
• RGB methods are computationally expensive, sensitive to lighting, and raise privacy issues.
• Single-photon LiDAR offers a privacy-preserving, robust alternative for 3D imaging.

Purpose of the Study:

• To develop a fast, long-range human activity recognition system using single-photon LiDAR and deep learning.
• To address data limitations by generating synthetic datasets.
• To create and release novel datasets for single-photon LiDAR-based activity recognition.

Main Methods:

• Combined an eye-safe single-photon LiDAR system with a deep learning pipeline.
• Generated synthetic datasets using motion capture data and 3D models.
• Trained a recurrent neural network on depth-image sequences of six activities.

Main Results:

• Achieved over 80% accuracy in human activity recognition, even in challenging scenarios.
• Demonstrated the system's capability for continuous and fast inference.
• Released two new long-range single-photon LiDAR datasets (325m and 1.4km).

Conclusions:

• Single-photon LiDAR combined with deep learning is effective for fast, long-range human activity recognition.
• The approach overcomes limitations of traditional RGB sensing.
• The developed system and datasets advance the field of privacy-preserving remote sensing.