Real-Time Fall Detection

• 

  Experimental setup for building fall dataset. Left: Floor plan of experimental region. Right: Fall event at the view of camera #4.

• 

  The architecture of MoViNets.

Inspiration

Accidental falls among the elderly without emergency medical assistance are a major cause of death, particularly for one person alone.

What it does

A feasible solution is to set up a fall detection system on the embedded system with the advantage of convenience and tiny size.

How I built it

We develop a real-time fall detection system for low computational resources using MoViNets. The proposed multi-view depth dataset is utilized to train a depth-based fall detection model.

Challenges I ran into

Although the embedded system is low-cost for home care or senior care facility, the limited computational resource makes it difficult to deploy a deep learning model to realize real-time detection.

Accomplishments that I'm proud of

The accuracy, specificity, and sensitivity reaches 99.21%, 99.69%, 96.77%, respectively. When the model is deployed to the embedded system, we use multi-processing to capture the frames and infer the input clips. The latency of the model is 208ms and the inference time interval of each clip is 1.6 sec when we deploy to a commercial embedded system (Nvidia Jetson Nano 2GB).

Built With

• depth-based-fall-dataset
• depth-camera
• jetson-nano
• python
• pytorch