/RepDetect

AI Android application crafted with Kotlin that harnesses the power of MediaPipe Pose Landmark Detection to deliver real-time feedback on exercise form while accurately counting repetitions

Primary LanguageKotlinApache License 2.0Apache-2.0

RepDetect

Make counting of exercise easier with RepDetect.

App Logo

Mobile project with AI

Welcome to RepDetect, an innovative Android application developed with Kotlin that leverages the capabilities of MediaPipe Pose Landmark Detection. RepDetect accurately counts repetitions, enhancing your workout experience.

Table of Contents

Overview

Developed as part of our third-year course at Metropolia University of Applied Sciences, RepDetect is a collaborative effort by a team of four students. Our primary objective was to create an innovative fitness application adhering to specific guidelines and meeting the course requirements.

Features

  • Pose Detection AI Model: Accurate detection of user poses to evaluate exercise form.
  • Pose Classification: Identify and classify different exercise poses for precise feedback.
  • Repetition Counting: The AI model accurately counts repetitions for each exercise, helping users track their progress
  • Real-Time Feedback: The app uses MediaPipe Pose Landmark Detection to analyze the user's exercise form in real-time. Currently it checks the confidence level of yoga and informs the user about the confidence percentage on it.
  • Exercise Plan: Users can create customized exercise plans, specifying the number of repetitions for each exercise. Plans can be based on specific exercises, allowing for a personalized workout routine.
  • Room Database: RepDetect utilizes the Room database to store and manage exercise plans and workout results.
  • Voice Notifications: Receive voice notifications at the start and completion of exercises for a seamless workout experience.
  • Exercise History: The app maintains a detailed exercise history, providing users with insights into their performance over time. Users can view weekly and daily summaries, helping them stay motivated and on track.
  • Camera Flipping: Easily switch between front and back cameras for varied workout perspectives.

Supported Exercises

  1. Pushup
  2. Squat
  3. Situp
  4. Deadlift
  5. Chest Press
  6. Shoulder Press
  7. Lunges
  8. Yoga Poses:
    • Warrior Yoga
    • Tree Yoga

AI Model Training Process

The training model is based on transfer learning and the training is done in Google Colab :

Step 1

  • Collect approximately 200 images for each exercise pose from various sources.
  • Compress images into a folder ("fitness_poses_images_in.zip") and upload to the specified section.

Step 2

  • Upload a sample video of a particular posture exercise for testing and model validation.

Step 3

  • Download the generated CSV file.
  • Divide the output file into smaller CSV files based on posture names (e.g., "squats.csv").
  • Integrate all files into the assets/pose directory of the Android application.

Getting Started

Prerequisites

  • Android Studio: Make sure you have the latest version of Android Studio installed.
  • Kotlin: The project is written in Kotlin, so familiarity with Kotlin is recommended.

Installation

  1. Clone the repository

     https://github.com/SurajKRB/RepDetect.git

  2. Open the project in Android Studio.

  3. Build and run the app on your Android device or emulator.

OR--

  1. Download the app-release.apk to Android phone
  2. Remember to "Allowing app installs from Unknown Sources in Android"

Usage

  1. Open the app on your Android device.

  2. Choose the exercise you want to perform (pushup, squat, situp, lunges, deadlift, chest press, shoulder press or yoga).

  3. Set up your exercise plan, specifying the number of repetitions for each exercise.

  4. Start your workout, and the app will count your repetitions.

  5. For yoga, it shows the confidence level. So it changes color according to confidence level. If it is above 90%, it shows green.

  6. View your exercise history to track your progress over time.

What was found

  • Challenges in Model Training:

Training the pose detection model posed significant challenges due to the requirement for a substantial number of images with perfect exercise form for each workout. The process demands meticulous curation of diverse training data to ensure the model's accuracy.

  • Successful Pose Detection and Repetition Count:

With proper training data, accurate pose detection and repetition counting for various exercises is possible.

  • Conflicts in Similar Exercises:

Similar forms of exercises can present conflicts for the pose detection model. Distinguishing between exercises with comparable poses requires careful consideration during both data collection and model training.

  • Pose Landmarker Model Details:

The pose landmarker model tracks 33 body landmark locations, representing the approximate position of various body parts.The arrangement of landmark points provides comprehensive information about body orientation, limb positions, and spatial relationships.

  • Real-Time Operation and Continuous Analysis:

MediaPipe Pose Landmark Detection operates in real-time, allowing continuous tracking and analysis of body movements during exercises.

  • Confidence Scores and Reliability:

The pose detection model typically generates a confidence score for each landmark. This score serves as a valuable metric to assess the reliability of detected landmarks.

  • Application of Detected Landmark Data:

The detected landmark data plays a crucial role in training machine learning models for broader analysis beyond simple pose detection.

What Worked

  • Efficiently utilized MediaPipe Pose Landmark Detection for accurate pose identification.
  • Segregated training data into individual CSV files, ensuring future scalability.
  • Successfully integrated trained models into the app for real-time feedback.
  • Processed data extracted from the detection result and presented meaningful information to the user.

What Didn't Work and Future Improvements

  • Challenges in capturing a wide range of exercise variations for diverse training data.
  • Pose landmarker model tracking 33 body landmarks affected classification for exercises with similar poses.
  • Yoga pose detection accuracy may need improvement, especially in providing feedback for pose correction.
  • Future improvement involves expanding the variety of exercises, requiring additional training and pose detection adjustments.

Pros and Cons of the App

Pros:

  • Accurate pose detection and repetition count
  • Customized exercise plans
  • Exercise history and performance tracking
  • Diverse supported exercises
  • Voice notifications and seamless workout experience
  • Gif images for learning how to do the consecutive exercise

Cons:

  • Challenge in model training
  • There might be conflicts in similar Exercises
  • The accuracy of yoga pose detection may need improvement

Screenshots

logoScreenhomeScreenexternalCameraScreengifworkoutScreenyogaresultScreenexerciseListScreenplanListScreen profileScreen

Demo Video

Click to watch

Watch the video

Libraries

AndroidX Libraries:

  • AppCompat Library - androidx.appcompat:appcompat:1.6.1
  • Material Design Library - com.google.android.material:material:1.10.0
  • ConstraintLayout Library - androidx.constraintlayout:constraintlayout:2.1.4
  • Play Services Vision Common - com.google.android.gms:play-services-vision-common:19.1.3
  • Camera Core Library - androidx.camera:camera-core:1.3.0
  • Google ML Kit Common - com.google.mlkit:common:18.9.0
  • Pose Detection Common - com.google.mlkit:pose-detection-common:17.0.0
  • Pose Detection Accurate - com.google.mlkit:pose-detection-accurate:17.0.0
  • Pose Detection - com.google.mlkit:pose-detection:17.0.0
  • Navigation Runtime - androidx.navigation:navigation-runtime-ktx:2.7.5
  • UI Automator - androidx.test.uiautomator:uiautomator:2.2.0
  • Room Database - androidx.room:room-runtime:2.6.0
  • Room KTX - androidx.room:room-ktx:2.6.0
  • Fragment Testing - androidx.fragment:fragment-testing:1.6.2
  • JUnit - junit:junit:4.13.2
  • Espresso Core - androidx.test.espresso:espresso-core:3.5.1
  • RecyclerView - androidx.recyclerview:recyclerview:1.3.2
  • RecyclerView Selection - androidx.recyclerview:recyclerview-selection:1.1.0
  • Kotlin Standard Library - org.jetbrains.kotlin:kotlin-stdlib-jdk7:1.5.31
  • Multidex Library - androidx.multidex:multidex:2.0.1
  • ViewPager2 - androidx.viewpager2:viewpager2:1.0.0
  • Dots Indicator - com.tbuonomo:dotsindicator:4.3
  • Curved Bottom Navigation - np.com.susanthapa:curved_bottom_navigation:0.6.5
  • Navigation Fragment KTX - androidx.navigation:navigation-fragment-ktx:2.7.5
  • Navigation UI KTX - androidx.navigation:navigation-ui-ktx:2.7.5
  • Fragment KTX - androidx.fragment:fragment-ktx:1.3.6

Pose Detection Libraries:

  • Pose Detection with Default Models - com.google.mlkit:pose-detection:18.0.0-beta3
  • Pose Detection with Accurate Models - com.google.mlkit:pose-detection-accurate:18.0.0-beta3
  • Google ML Kit Camera - com.google.mlkit:camera:16.0.0-beta3

Other Android Libraries:

  • Google Gson - com.google.code.gson:gson:2.8.6
  • Google Guava for Android - com.google.guava:guava:27.1-android
  • CameraX - androidx.camera:camera-camera2:1.0.0-SNAPSHOT, androidx.camera:camera-lifecycle:1.0.0-SNAPSHOT, androidx.camera:camera-view:1.0.0-SNAPSHOT
  • On Device Machine Learnings - com.google.android.odml:image:1.0.0-beta1
  • MPAndroidChart - com.github.PhilJay:MPAndroidChart:v3.1.0
  • GIF Drawable - pl.droidsonroids.gif:android-gif-drawable:1.2.23
  • Glide - com.github.bumptech.glide:glide:4.16.0 (compiler: com.github.bumptech.glide:compiler:4.16.0)

Testing Libraries:

  • AndroidX Core Library - androidx.test:core:1.4.0
  • AndroidJUnitRunner - androidx.test:runner:1.4.0
  • JUnit Rules - androidx.test:rules:1.4.0
  • Assertions - androidx.test.ext:junit:1.1.3

ViewModel and LiveData:

  • Lifecycle LiveData - androidx.lifecycle:lifecycle-livedata:2.3.1
  • Lifecycle ViewModel - androidx.lifecycle:lifecycle-viewmodel:2.3.1

License

Apache License 2.0 license

References

Pose landmark detection guide

Train the model

Contributors

Chi Nguyen

Giao Ngo

Suraj Rana Bhat

Shilpa Singh Yadav

(back to top)