AFD-CNN(ADL and Fall Detection Convolutional Neural Networks)
our paper[https://ieeexplore.ieee.org/document/8662651]
Sensor data to img
if the 3-axes of the human motion model are considered as the 3 channels of a RGB image, the value of the XYZ axial data can be mapped into the value of the RGB channel data in a RGB image respectively. Namely, each 3-axial data can be converted into an RGB pixel. The 400 pieces of 3-axial data cached in the sliding window can be viewed as a bitmap with size of 20 or 20 pixels.
you can use .utils.transform.data2image func to make sensor data to img
ADLs and fall data graph
sensor data to img
Net construct
we use imgs to train our network
Net performance
- accuracy = 0.978718
| Class | Sensitivity | Specificity |
|---|---|---|
| Fall | 1.000000 | 0.998654 |
| Walk | 0.969072 | 1.000000 |
| Jog | 0.983051 | 0.993243 |
| Jump | 0.948980 | 0.998684 |
| up stair | 0.989474 | 0.997379 |
| down stair | 0.967213 | 0.991848 |
| stand to sit | 0.981481 | 0.998667 |
| sit to stand | 0.990476 | 0.997344 |
| Average | 0.978718 | 0.996977 |
Requirenments
- python3
- tensorflow 1.4.0
- pandas
- numpy
- matplotlib
How to train and test
python ./src/cnn.py
Dataset
we need two public datasets.
-
1.MobiFall＆MobiAct DataSet http://www.bmi.teicrete.gr/index.php/research/mobiact
-
2.SisFall http://sistemic.udea.edu.co/en/investigacion/proyectos/english-falls/