Analytics


Fall Detection



We developed a hidden Markov model (HMM)-based fall detection system to detect falls automatically using a single motion sensor for real-life home monitoring scenarios.

Our system outperformed benchmark systems on the two datasets with maximum F-measures of 0.986 and 0.880, respectively (Yu & Chen, 2017).

Yu, S., & Chen, H. (2016, November). Fall Detection with Orientation Calibration Using a Single Motion Sensor. In International Conference on Wireless Mobile Communication and Healthcare (pp. 233-240). LNICST, Springer.


Different phases within a fall event derived from acceleration signals



Fall Risk Assessment



We developed a convolutional neural network (CNN)-based system to provide fall risk assessment based on motion sensor data in clinical settings.

Our model achieved F-measure of 0.942, outperforming benchmark systems by at least 12 percent (Yu & Chen, 2017).

Yu, S., & Chen, H. (2016, November). Fall Detection with Orientation Calibration Using a Single Motion Sensor. In International Conference on Wireless Mobile Communication and Healthcare (pp. 233-240). LNICST, Springer.


Our Proposed 2D Convolutional Neural Network Architecture



Activity of Daily Living (ADL) Recognition



We developed a novel 3-step deep learning architecture for ADL recognition. We introduced a CNN to extract interactions from raw sensor input. Interactions then combined with object information to constitute the gesture sequence as the input of the sequence-to-sequence (Seq2Seq) recurrent neural network (RNN) to recognize ADLs.

Our result outperformed the state-of-the-art model with weighted F1-measure 0.842 over 0.799. On average, our deep ADL recognition system can correctly label 240 seconds of the ADL sequences in a 5 minutes period (Zhu & Chen, 2017).

Zhu, H., & Chen, H. (2017). A Deep Learning Approach for Recognizing Activity of Daily Living (ADL) for Senior Care: Exploiting Interaction Dependencies and Temporal Patterns. Working Paper, Artificial Intelligence Lab, University of Arizona.


Recognizing ADLs from Raw Accelerometer Input

Seq2Seq Architecture with Gated Recurrent Units (GRU)


Our Publications


  • • Yu, S., & Chen, H. (2017). Motion Sensor-Based Fall Risk Assessment with 2D Heterogeneous Convolutional Neural Networks. Working Paper, Artificial Intelligence Lab, University of Arizona.
  • • Zhu, H., & Chen, H. (2017). A Deep Learning Approach for Recognizing Activity of Daily Living (ADL) for Senior Care: Exploiting Interaction Dependencies and Temporal Patterns. Working Paper, Artificial Intelligence Lab, University of Arizona.
  • • Yu, S., & Chen, H. (2016, November). Fall Detection with Orientation Calibration Using a Single Motion Sensor. In International Conference on Wireless Mobile Communication and Healthcare (pp. 233-240). LNICST, Springer.
  • • Maimoon, L., Chuang, J., Zhu, H., Yu, S., Peng, K. S., Prayakarao, R., Chen, H., et al. (2016, December). SilverLink: Developing an International Smart and Connected Home Monitoring System for Senior Care. In International Conference on Smart Health (pp. 65-77). LNCS, Springer.
  • • Chuang, J., Maimoon, L., Yu, S., Zhu, H., Nybroe, C., Hsiao, O., Chen, H., et al. (2015, November). SilverLink: Smart Home Health Monitoring for Senior Care. In International Conference on Smart Health (pp. 3-14). LNCS, Springer.