Time-Series Analysis for Fall Detection
This project aims to detect falls from time-series data collected from a wearable device containing an accelerometer. The primary goal is to design a model capable of accurately detecting falls from the given time-series data.
The data was gathered from a single subject and exclusively tested on the subject. It comprises three different activities: walking, falling, and kneeling.
To ensure clean data, noise was manually removed. Labels were added to indicate the activity types correctly. Additionally, during the preprocessing phase, the derivative of the movement across all axes was considered for improved results.
The model design consists of a Bi-directional LSTM to analyze the movement sequence across all axes, followed by an MLP to classify the activity type based on the captured sequence.
Due to the limited data with only one subject, the model exhibited signs of overfitting to the subject.
Dynamic range quantization was employed to reduce the model size, and conversion to tflite format enabled the model's deployment on mobile devices.