In this project I have trained a convolutional neural network to classify the state of a smartphone into the following states: walk, bike, rail travel, car drive. To create a data set I have programmed an Android app which records the accelerometer and the gyroscope while doing this activities. Later on I also integrated the trained model into the app to make live predictions on the smartphone. I did this project in spring 2018 during my free time besides my computer science studies. Here you can see a screenshot from the app:
To train the neural network I collected roughly 1000 samples from each class. After 15 epochs I have an accuracy of about 98.44% on previously unseen data. You can see the process of training in the following graphs:
Number of classes: 4
Class 0: 1047
Class 1: 1080
Class 2: 1020
Class 3: 1058
Train data: 3365
Val data: 632
Test data: 200
Could not find old network weights
Epoch 0, validation accuracy 0.6516
Epoch 1, validation accuracy 0.7078
Epoch 2, validation accuracy 0.8016
Epoch 3, validation accuracy 0.9359
Epoch 4, validation accuracy 0.8797
Epoch 5, validation accuracy 0.8906
Epoch 6, validation accuracy 0.875
Epoch 7, validation accuracy 0.9203
Epoch 8, validation accuracy 0.9828
Epoch 9, validation accuracy 0.975
Epoch 10, validation accuracy 0.9781
Epoch 11, validation accuracy 0.9781
Epoch 12, validation accuracy 0.9781
Epoch 13, validation accuracy 0.9797
Epoch 14, validation accuracy 0.9797
Final test accuracy 0.984375