SmartFall Dataset Description

• all fall data samples are equally 25 points
• the cycle repeats with 'ADL-FALL-ADL-FALL'

SmartFall LSTM & GRU Classification

• To make input dataset, I have partioned data into 40 points
• If all 25 points of fall data is included in partioned 40 points, I labelled it as 'FALL'
• Unless, it's all labelled as 'ADL'
• Since there is huge data imbalance between Fall and ADL, I shrinked ADL data portion as same as Fall data
• I also implemented cyclic learning rate

SmartFall LSTM & GRU Classification Results

	Precision	Recall	F1 Score	Accuracy
SmartFall LSTM	0.9963	0.8411	0.9121	0.9189
SmartFall GRU	0.9963	0.8442	0.9139	0.9205

MobiAct Dataset

• To compare the result between MobiAct dataset and SmartFall dataset, I resampled MobiAct dataset similar to SmartFall dataset
• To make input data, I have resampled all MobiAct Fall data's fall parts to 30 points and added 10 samples of ADL at the front and the end of fall samples
• Then I used partioned data with window size of 40(same as SmartFall data window size) to make input dataset
• You can find MobiAct Dataset at below url
• https://bmi.hmu.gr/the-mobifall-and-mobiact-datasets-2/
• You can see partioning code throught MobiAct_DataParsing.ipynb

Transfer Learning using MobiAct Dataset

• I have used pretrained model using SmartFall Dataset to MobiAct Dataset but because of the difference of data collected domain performance was really bad
• The chart below describe the result of training SmartFall data & testing on SmartFall data, using pretrained model to test on MobiAct data, relearning pretrained model using MobiAct data

• The chart shows that relearning through pretrained model works well