[home] [Personal Program] [Help]

---

Session: Motion
Session starts: Thursday 26 January, 14:30
Presentation starts: 15:30
Room: Room 531

---

Reda El Hail (KU Leuven, Dept of Computer Science, DTAI, Leuven.AI, B-2440 Geel, Belgium ; Flanders Make, DTAI-FET, Belgium)
Dominique Schreurs (KU Leuven, Dept of Electrical Engineering, ESAT-WAVECORE, B-3000 Leuven, Belgium)
Peter Karsmakers (KU Leuven, Dept of Computer Science, DTAI, Leuven.AI, B-2440 Geel, Belgium ; Flanders Make, DTAI-FET, Belgium)

Abstract:
Healthcare systems encounter many difficulties due to a shorthand of hospital staff members. For instance, in clinics some patient rooms need constant monitoring which evidently cannot be performed by caregivers. Both wearable and contactless technological solutions have been proposed. Recent works have considered radar sensors since such sensors do not need a patient to wear anything, are robust to different lighting conditions and preserve the patients’ privacy. In this work a frequency modulated continuous waveform radar along with signal processing and machine learning algorithms is assessed for the purpose of recognizing human movement events. Data was collected from 10 people using two radars at two positions (ceiling, wall), and a camera for annotation purposes. The dataset contains 246 “walk” events, 96 “sit” events, 96 “stand-up” events, 42 “hand movement” events, 108 “lay dawn” events and 108 “get up” events. A radar configuration was carefully chosen based on the desired maximum and resolution of range and velocity. Raw signals captured by the radar were processed using a standard processing pipeline into Doppler-time maps. To detect low-level movement events from the Doppler-time maps two deep learning model architectures, a Convolutional Neural Network (CNN) and a Convolutional Long-Short Term Memory (ConvLSTM), were assessed and compared to each other. Experiments were carried out to determine appropriate hyper-parameter values such as the time horizon for the Doppler- time maps. It was observed that the additional temporal modeling that ConvLSTM provides did not improve performance. The best CNN model achieved 85.5% average recall in a leave-one- person-out experiment. When inspecting the model misclassification in more detail it turned out that there are primarily two explanations for most misclassifications. The first concerns quick events, like sitting, which take less than a second to complete, and the second involves unusual data from volunteers who were moving more slowly compared to others. Furthermore, the radar mounted on the wall had a slightly higher performance in terms of classification accuracy compared to the radar on the ceiling. Based on these preliminary findings, future research will design new models that can deal better with both short and longer events and focus on relative Doppler speed patterns.