TrustFall

• 

  Front page of app

• 

  Heatmap of past falls

• 

  Planar undistortion

• 

  Floor plan of house with coordinates of fallen elderly

What it does

TrustFall is an application mainly aimed at elderly staying in rental hdbs and nursing homes. Using a Computer Vision AI model we trained using transfer learning, our application is able to accurately detect falls through cctv footages. An alarm would then be triggered, notifying volunteer centers such as Lion Befrienders and Red Cross. This reduces the response time taken, enabling elderly to receive timely and adequate assistance as volunteers pass by their homes. Through past fall cases, heatmaps are also produced for elderly to be aware of danger or high risk zones around their homes, thus helping them proactively mitigate falls. Also, such heatmaps are useful for strategic camera placements to safeguard elderly in the best way.

How we built it

We first used almost over 3000 images augmented using several means such as image distortion (to produce accurate images regardless of camera angle) and shearing etc. This dataset was then split in 70-20-10 fashion to train, validate and test our model. As an individual component our YoloV5 model excelled with MaP@50 rating at 0.98. Next, we incorporated the CV Model into the TrustFall application coded using Flask API & HTML. Using a planar image undistortion software that was produced with some of our own math and online materials, we managed to obtain a top down view of the flooring in a house which can be overlayed on a rough floor plan to identify exact coordinates of locations within the home. This can be used by volunteers who make use of our application to better navigate to the elderly especially if its a large apartment. Using some data from our past Project, we managed to form a Heat Map to identify vulnerable elderly fall locations so that elderly can take proactive measures against falls.

Built With

• computervision
• css
• flask
• heatmap
• html
• planar-undistortion
• python
• sqlite