Prediction model for human activity recognition

Recognise a human activity from a wide range of possiblities using a Deep Learning Neural Network

Goal:

To recognize variety and wide range of activities performed by humans, using data collected from a combination of sensors from smartphone and smartwatch devices.

Implementation:

Have implemented deep learning concepts and developed two models with Feed Forward Fully Connected Network and Recurrent Neural Network adaptive frameworks. Fuse the inputs from two sensors – accelerometer and gyroscope of smartphone and smartwatch, which is in turn feed in to train the models. Later we used cross visualization technique to evaluate the efficiency of the models

Accuracy and hyperparameters tuning:

Accuracy of FNN model - 56.22% of model accuracy by tuning with the below hyper parameters. Recurrent Cycle - 1 Neurons – 10 Epochs – 50 Learning Rate – 0.01

Model Performance and results:

Model Accuracy

Model Loss

Confusion Matrix FNN

Confusion Matrix RNN

Please refer to the research paper(unpublished) Heterogeneous Human Activity Recognition with Dynamic Sensor Fusion using Deep Learning Models.pdf for futher explanation.

Contributions:

1. Bhuvaneshwaran Ravi
2. Serlin Tamilselvam
3. Jayashree Srinivasan