Implement 2 frame lookback to determine fall instead of 1

[ivelin]

Is your feature request related to a problem? Please describe.

Currently with look back 1 image frame with 1 second spacing between images to detect a fall.

This research study concludes that falls last anywhere between 600 and 800 ms. We currently expect 1 second gap between frames, which seems to be a reasonable choice that should be long enough to capture an image before and after a fall. The question remains open how to handle sequence of frames that may come in (before and middle) of a fall or (middle and after) a fall. In these cases the fall may be incomplete between the two comparison frames and we won't detect it with the current algorithm. Something to think about. May we look 2 frames back instead of 1, which ensures that a fall is certainly captured either in the (1,2), (2,3) or (1,3) pair of frames.

Describe the solution you'd like

Let's assume image[t] is the camera image input at time t.

Currently with only 1 image lookback with 1 second spacing between images and a fall taking between 600-800ms, we have these possible cases:

[image[t-1], fall_start, fall_end, image[t]] - > detection! (body vector angle between image[t] and image[t-1] should be close to 90', certainly over 60')
[image[t-1], fall_start, image[t], fall_end] - > no detection (body vector angle between image1 and image2 could be less than 60')
[ fall_start, image[t-1], fall_end, image[t]] - > no detection (body vector angle between image1 and image2 could be less than 60')

What if we look not only 1 but 2 images back with 1 second spacing between each:
[image[t-2], image[t-1], fall_start, fall_end, image[t]] - > detection! (body vector angle between image[t] and image[t-1] should be close to 90', certainly over 60')
[image[t-2], fall_start, image[t-1], fall_end, image[t]] - > detection (body vector angle between image[t-2] and image[t] should be close to 90', certainly over 60')

It seems like with 3 frames (current time t, t-1 and t-2) lookback we should be able to capture any true fall as long as these assumptions hold true:

• Image frames are spaced no less than 1 second apart and no more than 2 seconds apart.
• A fall takes 600-800ms from initiation to ground contact
• A person who fell down stays on the ground for at least 2 seconds. If they jump up faster than that, it's probably not a case to be alarmed about.
• The angle between the body vectors of a standing and a fallen person is at least 60' (configurable).

With the assumptions above, a true will be detected either between [image[t-1], image[t]] or between [image[t-2], image[t].

Additional context
A different research study on ambient health monitoring discovered that only 2% of video frames are needed to accurately detect a person activity. 2% at 60fps is 1.2fps. This is consistent with the 600-800ms finding of the study above. Although Dr. Fei's study relies on RNNs to find which 2% are the ones required for the activity detection, it is still helpful to educate our simpler heuristic formula.

[github-actions]

🎉 This issue has been resolved in version 1.13.0 🎉

The release is available on GitHub release

Your semantic-release bot 📦🚀