Human Activity Recognition and Fall Detection for Elderly

About the project

Many senior citizens stay alone for most of the day, and after a fall they may lie on the floor for hours before anyone knows. This fall detection Android project solves that with the phone they already carry. CareStep reads the accelerometer and gyroscope 50 times a second, and a small TensorFlow Lite model on the phone recognises six activities: walking, stairs, standing, sitting, lying and fall. When a fall is detected, a 30-second "Are you okay?" countdown starts. If the elder does not answer, the linked family members get an alert with the location and an SMS goes to the emergency contacts.

It is more than an emergency button. CareStep also tracks daily activity and steps, reminds medicines, tells the family about missed doses, collects a daily wellness check-in and runs a fall-risk self-test.

How it works

• The elder logs in and monitoring starts as a foreground service (works with the screen off)
• Every 2.56-second sensor window becomes 16 features and the TFLite model gives the activity
• Minutes of each activity and steps are synced to the server every 5 minutes
• A fall (model probability or strong impact + lying still) opens a 30-second countdown
• No answer → alert with location saved, family notified, SMS with map link sent
• The family attends the alert, checks medicines adherence and downloads a weekly PDF
• Admin manages users, links, alerts, questions, notices, reports and detection settings

Users

• Elderly User: live activity, fall countdown, SOS, medicines, daily check-in, fall-risk self-test, activity reports and PDF
• Family Member: link elder with a code, dashboard, alerts with map and call, medicine adherence, assign reminders, weekly PDF report
• Admin: dashboard with charts, elders, families, link approval, alerts, logs, test questions, notices, printable reports, settings

Highlights

• 3 roles, 22 app screens, 13 admin pages, 35 APIs, 14 database tables
• Android activity recognition with an on-device TFLite model: 97.00% test accuracy, 98.75% fall recall
• Fall-risk score from 12 STEADI-style questions + 30-second chair-stand test counted by the accelerometer
• Medicine reminders with Taken button, missed-dose alerts to family and a 7-day adherence grid
• Validations in the app and on the server (email, 10-digit mobile, password, dates, duplicates)
• Secure: password hashing, prepared statements, role and link checks, CSRF tokens in admin forms
• Ready sample data and demo logins; demo buttons to simulate walk, sit, lie and fall on the emulator

What you get

• Complete Android Studio source code (Java) and PHP admin panel source code
• MySQL database file with sample data
• ML folder: dataset, training script, TFLite model and accuracy report
• Project report (Word, 50+ pages) with all UML, DFD and ER diagrams
• 2-page project abstract, project explanation with 20 viva questions
• Installation / setup guide (PDF) and setup help

Who can use this project

Final year BE / B.Tech (Computer, IT, ENTC), MCA, BCA, M.Tech and Diploma students looking for an Android project with source code in machine learning, healthcare or elderly care. It suits students who want a fall detection app project or a human activity recognition project with full documentation and report.

Frequently asked questions

Does it need a smartwatch or any extra device? No. It uses only the phone's accelerometer, gyroscope and step counter.

How do I show a fall in the college demo? The Live Activity screen has Simulate buttons for walk, sit, lie and fall, so the full countdown and alert flow works even on the emulator.

Does the model need internet? No. The TFLite model runs on the phone. Internet is needed only to sync data and for the optional Gemini summary.

Can I retrain the model with real data? Yes. The ML folder has the training script; put real recordings (for example UCI HAR or SisFall) in the same format and run it again.