Human Being Detection from UWB NLOS Signals: Accuracy and Generality of Advanced Machine Learning Models.

Area of Science:

• Electrical Engineering
• Computer Science
• Robotics

Background:

• Detecting humans in non-line-of-sight (NLOS) scenarios is crucial for safety and security applications.
• Traditional methods often struggle with environmental variability and unknown observation models.

Purpose of the Study:

• To evaluate the effectiveness of machine learning (ML) techniques for human detection in NLOS conditions using ultra-wideband (UWB) radar.
• To compare various ML approaches and input representations for accuracy and adaptability in realistic environments.

Main Methods:

• Conducted extensive UWB radar measurements in diverse realistic environments.
• Investigated scenarios with different body orientations, obstacle materials, and radar positions (mobile cart, handheld).
• Analyzed supervised and unsupervised, symbolic and non-symbolic ML methods.

Main Results:

• ML techniques demonstrated high accuracy and adaptability in detecting NLOS human beings.
• ML methods showed effective knowledge transference, generalizing to unseen cases.
• Modern ML approaches outperformed traditional techniques in flexibility and overcoming data limitations.

Conclusions:

• Machine learning provides a flexible and effective solution for NLOS human detection with UWB radar.
• ML's ability to generalize surpasses traditional methods, especially in challenging or data-scarce situations.
• This approach eliminates the need for environment-specific configurations, enhancing practical applicability.