[FEATURE] Detect anomalies in daily routines

[ivelin]

Most people and especially elderly have a recurring daily pattern of movement around their home.

An AI sequence model could learn to predict movements and detect anomalies. Examples:

• A person did not get out of bed all morning.
• A person is spending much longer time in the bath than usual.
• A person did not do any stretches and exercises all day.
• A person is lying on the couch in front of the TV past their typical bedtime.
• A person did not open the fridge all day.
• A person did not visit bathroom all day.

[ivelin]

After gathering additional field experience, the following ideas emerged:

Context: Falls occur often by the bedside in the bedroom and in bathroom. Both of these are private, space constrained areas with suboptimal visibility. Makes it hard to gather good training data for fall detection.

Possible solutions:

1. Short time window (10-15 second) time series that capture a sequence of pose detection keypoints and feed them to an anomaly detection model for acute problems such as falls.

   • Anomaly detection with TensorFlow and Autoencoders https://www.youtube.com/watch?v=2K3ScZp1dXQ
   • This might be practical for detecting anomalies in basic daily routines such as getting out of bed, walking into/out of bathroom , opening fridge, walking into/out of closet.

2. Longer time window (24 hour) that compares time series from the most recent 24 hour and the historical 24 hour data distribution to detect anomalies and predict potential upcoming complications that are signaled by change in daily routines.

   • This model can feed from the short time window model and combine into single activity detection data points.
   • This model might also combine inference data from multiple ambianic boxes.
   • The resulting time series might look like this:
   • around 8am getting out of bed
   • 9:30am - 10am open fridge
   • 11:30am - 12:00pm visit restroom
   • around 8pm bedtime

3. Experiment with even longer term time windows to learn about weekly, monthly, seasonal, annual cycles that might relate to health and wellbeing.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.