estellekao / ee292d_project Goto Github PK

Falls represent a considerable medical danger to the world's population. Many fall detection systems have been implemented to rectify this issue, such as the Apple Watch fall detection, Life Alert, and various smartphone IMU applications. However, the models implemented in these devices are either inference-only or cloud-based. Given the diversity of lifestyles and living conditions in the world, a model that can retrain from data it collects on the edge could provide users with a more robust fall prediction system. This report presents an experimental device setup and data collection to test such an adaptive fall detection model on the edge. The device is initially trained on data from the MobiAct dataset and then retrained on the edge using data collected from an LSM330DLC Accelerometer/Gyroscope [1]. Our neural network-based device retrained on the edge achieved an accuracy of 71.0% and 80.0% on two activities we taught our model to label as a non-fall with only 1 and 3 retraining events respectively. We also achieved an accuracy 94.0% on one new activity we taught the model to consider a fall with 4 retraining events.

The MLP model is adapted from ttshiz's git repo.

Addition set up instructions are descrbied in Micropython/README.md

Addition set up instructions are descrbied in UpBoard/README.md

[1] Chatzaki C., Pediaditis M., Vavoulas G., Tsiknakis M. (2017) Human Daily Activity and Fall Recognition Using a Smartphone’s Acceleration Sensor. In: Rocker C., O’Donoghue J., Ziefle M., Helfert M., Molloy W. (eds) Information and Communication Technologies for Ageing Well and e-Health. ICT4AWE 2016. Communications in Computer and Information Science, vol 736, pp 100-118. Springer, Cham, DOI 10.1007/978-3-319-62704-5_7.

[2] Shizume, T., & Salem, W. (n.d.). TTSHIZ/CS539_FALL_PREDICTION. GitHub. Retrieved December 14, 2022, from https://github.com/ttshiz/CS539_Fall_Prediction