Machine Learning-Based Human Detection Using Active Non-Line-of-Sight Laser Sensing.

Area of Science:

• Photonics and Sensing
• Machine Learning for Security
• Search and Rescue Technologies

Background:

• Active non-line-of-sight (NLOS) human detection utilizes indirect photon reflections to identify hidden individuals.
• Simulated post-disaster rubble environments were used to test sensing capabilities.

Purpose of the Study:

• To evaluate the effectiveness of a single-photon avalanche diode (SPAD) and time-correlated single-photon counting (TCSPC) system for NLOS human detection.
• To compare the performance of machine learning models in classifying human presence based on time-photon waveforms.

Main Methods:

• A SPAD and TCSPC system acquired time-photon waveforms in controlled NLOS environments.
• Data preprocessing involved normalization, histogram shaping, and signal windowing.
• Convolutional Neural Network, Gated Recurrent Unit, and Random Forest models were trained for classification.

Main Results:

• All models achieved 100% sensitivity in detecting human presence.
• Random Forest exhibited the highest overall accuracy and specificity, particularly in human-absent scenarios.
• Tree-based classifiers demonstrated superior robustness to statistical variations under limited photon conditions.

Conclusions:

• Low-cost SPAD-based NLOS sensing systems offer reliable human detection in indirect-observation scenarios.
• Random Forest models are effective for classifying human presence using time-photon histogram data.
• The study highlights the potential of photon-based sensing for search and rescue operations.