Fall Detection

Project Description

Fall Detection is an Arduino based system that detects whether a person has fallen. It is helpful for older people who might fall and need immediate care and assistance. Our system detects the fall and can be used to alert the authorities. It is based on M.Saleh et.al's implementation in [1]. We thank M. Saleh et. al for sharing their data and code in the public domain.

Tools Required

1. Google Colab
2. Arduino IDE
3. Arduino Nano 33 BLE Sense

Dataset

The Jupyter Notebook utilizes the FallAllD Dataset publically available on IEEE DataPort. The notebook processes the useful data to generate a meaningful dataset. Original Dataset contains data from IMU and Barometer(b) however the notebook extracts only Accelerometer(a) and Gyroscope(g) Data as the remaining data appears to be meaningless. It can be seen from plotting the heatmap of the data.

It can be clearly seen that barometer shows almost no changes when a person falls. We also know that magnetometer(m) gives us our heading from North and therefore doesn't carry enough meaning to be determining if a person has fallen. Therefore we have utilized only Accelerometer and Magnetometer data only. We have further downsampled the data to 50 Hz to reduce the overall computation cost. Otherwise we would have to increase the depth of our Neural Network.

Model Architecture

We have utilized a Convolutional Neural Network to predict the falls. The model is designed with following goals:

1. Reduce the Model Size to fit in our device.
2. Reduce the computation cost to make our predictions quicker.
3. Provide reliable predictions to detect falls.

Model Optimizations

We convert our model to Tensorflow Lite and quantize our model by providing it a representative dataset. The quantized model is around 140 KB in size.

converter=tf.lite.TFLiteConverter.from_keras_model(model)
tflite_no_quant=converter.convert()
converter.representative_dataset=representative_ds_gen
converter.optimizations=[tf.lite.Optimize.DEFAULT]
tflite_model=converter.convert()

We then convert our model into a FlatBuffer to fit it in our device and interpret it using a Tensorflow Lite interpreter.

# Install xxd if it is not available
!apt-get update && apt-get -qq install xxd
# Convert to a C source file, i.e, a TensorFlow Lite for Microcontrollers model
!xxd -i {MODEL_TFLITE} > {MODEL_TFLITE_MICRO}
# Update variable names
REPLACE_TEXT = MODEL_TFLITE.replace('/', '_').replace('.', '_')
!sed -i 's/'{REPLACE_TEXT}'/g_model/g' {MODEL_TFLITE_MICRO}

References

1. M. Saleh, M. Abbas and R. B. Le Jeannès, "FallAllD: An Open Dataset of Human Falls and Activities of Daily Living for Classical and Deep Learning Applications," in IEEE Sensors Journal, vol. 21, no. 2, pp. 1849-1858, 15 Jan.15, 2021, doi: 10.1109/JSEN.2020.3018335.