Practical Implementation of Robust Human Activity Cataloguing with mmWave Radar

• Published in: 2024 IEEE Opportunity Research Scholars Symposium (ORSS) pp. 37 - 40
• Main Authors: Mii, Hiroki, Shah, Swarna, Lu, Bozhou, Liu, Alan, Lin, Yu-Tai, Sundaresan, Karthikeyan
• Format: Conference Proceeding
• Language: English
• Published: IEEE 15-04-2024
• Subjects: Accuracy; Computational modeling; Data models; Millimeter wave communication; Millimeter wave radar; Pipelines; Protection; Real-time systems; Robustness; Sensors
• DOI: 10.1109/ORSS62274.2024.10697944

Millimeter wave (mmWave) radar has recently become a popular choice of sensing technology for privacy-protecting health monitoring solutions. Though mmWave radar improves upon invasive alternatives like computer vision, the cost of privacy protection is sparser data. Many machine learning models designed to process this data are either computationally heavy or lack robustness and are impractical to deploy in real-time as required by applications such as physical therapy assistance, remote healthcare, and personal fitness assistance. Therefore, we aim to deploy a well-designed RF-driven Human Activity Cataloguer (RF-HAC) [1] for real-time operation. The model detects, categorizes, and counts repetitions of human exercises, and has a robust, accurate classification algorithm. To deploy this model into a real-time system, we focused on optimizing the data flow pipeline to stream data from the radar and process it to receive an output within seconds of the activity occurrence. Our system reduces the processing time from 3.1819 seconds to 1.1161 seconds, and our implemented data streaming pipeline produces a real-time output while maintaining a classification accuracy of 94%.