Data Scientist | Computer Scientist
- About Mediapipe
- Use Case
- Challenges
- Process
- Result
- Demonstrate Pose Estimation
- Machine Learning
- Conclusions
- Further Study
Mediapipe utilizes BlazePose, a lightweight convolutional neural network architecture for real-time human pose estimation on computer or mobile devices.
- Activity recognition
- Motion capture and augmented reality
- Training robots
- Motion tracking for consoles
Human pose estimation faces challenges such as dynamic changes in appearance, occlusion, and background variations, making it difficult for image processing models to identify fine-grained joint coordinates.
Pose estimation predicts and tracks spatial positions of a body or object by identifying key points such as joints. Two types include single and multi-pose estimation.
BlazePose landmark model predicts 33 pose landmarks.
Predicted body joints count exercise repetitions and provide audio feedback.
Real-time webcam data used to train models (Logistic Regression, Ridge Classifier, Random Forest, Gradient Boosting) for emotion prediction (Happy, Sad, Yay).
Various machine learning algorithms implemented to classify real-time emotions with high accuracy:
- Logistic Regression: Accuracy = 1.0
- Ridge Classifier: Accuracy = 1.0
- Random Forest Algorithm: Accuracy = 1.0
- Gradient Boosting Classifier: Accuracy = 0.998
The tool enables users to track workout repetitions automatically, improving convenience during exercise routines.
- Expand exercise repertoire to include yoga poses.
- Integrate additional features like face recognition for identifying users.
- Deploy emotion prediction model using AWS.
