Actify - The Activity Recognition App

This project applies machine learning techniques on data collected from a Respeck inertial measurement unit (IMU) to classify a range of activities and social signals, displaying the results in real-time in an Android app. The Respeck device is worn on the lower left ribcage, and collects linear acceleration data along three axes. Two Convolutional Neural Networks (CNNs) were trained on this collected data, one for classifying activities and one for classifying social signals. An overview of the application structure is shown below. Data is streamed from the Respeck device at 25 Hz, filling two buffers. Once the buffers are full, the two CNNs both make independent classifications, and then the oldest 50 data points are removed from the buffers. The classification results are displayed in the app and stored in a database so the user can reflect on their daily activity and social signal patterns.

The below tables show the activities and social signals to be classified along with their respective accuracies. The accuracies were measured as the average recall for each class in 5-fold cross validation. The "Other" social signal class includes laughing, talking, eating and singing.

Activity	Accuracy
Ascending stairs	96.13
Descending stairs	96.43
Sitting/Standing	99.06
Shuffle walking	94.43
Miscellaneous movements	91.50
Normal walking	97.83
Lying on back	97.17
Lying on left	96.12
Lying on right	96.96
Lying on stomach	98.98
Running	99.75

Social Signal	Accuracy
Breathing normal	86.83
Coughing	91.10
Hyperventilating	81.37
Other	90.41

Application Structure

The application is designed with a modular structure. Each module is responsible for a specific functionality. For example, ClassificationService in the services directory performs background classification of activities and social signals using machine learning models. The important modules in the application and their functions are shown in the table below.

Layer	Module Name	Function
Presentation Layer	MainActivity	Central hub for navigation and app initialization.
	LiveDataActivity	Real-time sensor data visualization and displaying classification results.
	HistoricalActivity	Displays historical activity and social signal data for user-selected dates.
Service Layer	ClassificationService	Performs background classification of activities and social signals using machine learning models.
	BluetoothSpeckService	Manages Bluetooth connectivity with the Respeck device.
	RespeckPacketDecoder	Decodes raw Respeck data packets into usable accelerometer data.
	RespeckPacketHandler	Processes decoded data for use in visualization and classification.
Data Layer	AppDatabase	Implements a Room database for persisting historical classification results.
	ActivityLogDao	Provides methods for managing activity classification records in the database.
	SocialSignalLogDao	Provides methods for managing social signal classification records in the database.
	DailyActivityLog	Primary table for storing daily activity classification records.
	DailySocialSignalLog	Primary table for storing daily social signal classification records.
	DataCleanupWorker	Periodically deletes old database entries to optimize storage.

Installing the App

The Actify app can be installed using Android Studio. Connect the mobile device to your computer via USB, ensure developer mode is enabled on the mobile device, and deploy the app directly from Android Studio. Pair the Respeck sensor with the mobile device via Bluetooth before starting the app. Once connected, the app will begin collecting data from the Respeck.