Human Falling Detection and Tracking

Using Tiny-YOLO oneclass to detect each person in the frame and use AlphaPose to get skeleton-pose and then use ST-GCN model to predict action from every 30 frames of each person tracks.

Which now support 7 actions: Standing, Walking, Sitting, Lying Down, Stand up, Sit down, Fall Down.

Prerequisites

Python > 3.6
Pytorch > 1.3.1

Original test run on: i7-8750H CPU @ 2.20GHz x12, GeForce RTX 2070 8GB, CUDA 10.2

Data

This project has trained a new Tiny-YOLO oneclass model to detect only person objects and to reducing model size. Train with rotation augmented COCO person keypoints dataset for more robust person detection in a variant of angle pose.

For actions recognition used data from Le2i Fall detection Dataset (Coffee room, Home) extract skeleton-pose by AlphaPose and labeled each action frames by hand for training ST-GCN model.

Pre-Trained Models

Tiny-YOLO oneclass - .pth, .cfg
SPPE FastPose (AlphaPose) - resnet101, resnet50
ST-GCN action recognition - tsstg

Basic Use

Download all pre-trained models into ./Models folder.
Run main.py

    python main.py ${video file or camera source}

Reference