Activity Recognition from Single Chest-Mounted Accelerometer:

Usage Guide:

File Descriptions:

• 'exploratory_data_analysis.py' :

  contains the code to perform basic Exploratory Data Analysis(EDA) using various Visualizations.

• 'train.py' :

  contains the code to train K-Nearest-Neighbors, XGBoost, and Neural Network Models.

• 'predict.py' :

  contains the code to load the pre-trained models and predict Activity using 3-DOF sensor raw values (X-acceleration, Y-acceleration, and Z-acceleration) as input.

• 'knn_model.pkl' :

  contains the KNN model that I have trained.

• 'xg_model.pkl' :

  contains the XGBoost model that I have trained.

• 'neural_network_model.h5' :

  contains the Neural Network model that I have trained.

• 'requirements.txt' :

  contains the packages and their exact versions required to run the python scripts in this repository.

Running Files:

• To perform EDA, use:

 python .\exploratory_data_analysis.py --directory_path 'C:\Users\mehra\PycharmProjects\VicaraAssignment\Activity Recognition from Single Chest-Mounted Accelerometer\Activity Recognition from Single Chest-Mounted Accelerometer'

Change the value of the directory_path to specify the directory where you have saved the data files.

• To train models, use:

python train.py --directory_path 'C:/Users/mehra/PycharmProjects/VicaraAssignment/Activity Recognition from Single Chest-Mounted Accelerometer/Activity Recognition from Single Chest-Mounted Accelerometer' --model_name 'knn'

Change the value of the directory_path to specify the directory where you have saved the data files. Change the value of the model name to train KNN, XGBoost, and NN according to your requirements.

• To predict, use:

python predict.py --model_path C:\Users\mehra\PycharmProjects\VicaraAssignment\neural_network_model.h5 --model_name 'neural_network' --x_acc 2145 --y_acc 2336 --z_acc 1947

Change the value of the directory_path to specify the directory where you have saved the data files. Change the value of the model name to train KNN, XGBoost, and NN according to your requirements. Change the values of X,Y, and Z acceleration according to your requirements.

Exploratory Data Analysis(EDA):

Activity Counts:

Cross-Tab:

Acceleration Trends For Each Activity:

Results:

K-Nearest-Neighbors Model Results:

Accuracy is 65.03243551818983 % for k value : 2

Accuracy is 69.79033681042088 % for k value : 3

Accuracy is 71.14899579635684 % for k value : 4

Accuracy is 72.19264102963308 % for k value : 5

Accuracy is 72.68540142197311 % for k value : 6

Accuracy is 73.25730447869635 % for k value : 7

Accuracy is 73.53599045098345 % for k value : 8

Accuracy is 73.86657325237428 % for k value : 9

Accuracy is 73.99605584098812 % for k value : 10

Accuracy is 74.21168716591417 % for k value : 11

Accuracy is 74.30691784731953 % for k value : 12

Accuracy is 74.39851575068764 % for k value : 13

Accuracy is 74.45819710415694 % for k value : 14

Accuracy is 74.59286937568115 % for k value : 15

Accuracy is 74.65021537184077 % for k value : 16

Accuracy is 74.70107426436245 % for k value : 17

Accuracy is 74.73843998131714 % for k value : 18

Accuracy is 74.75971768124967 % for k value : 19

Accuracy is 74.79500752504022 % for k value : 20

Accuracy is 74.82458871762935 % for k value : 21

Accuracy is 74.85183455290881 % for k value : 22

Accuracy is 74.93720483678447 % for k value : 23

Accuracy is 74.92241424048991 % for k value : 24

Accuracy is 74.992993928071 % for k value : 25

The optimum number of neighbors for this dataset is 25

KNN Accuracy Score for Train Set: 0.7627958451290807

KNN Accuracy Score for Test Set: 0.7509030048264052

Classification Report:

           precision    recall  f1-score   support

       0       0.41      0.03      0.06       811
       1       0.86      0.91      0.89    121138
       2       0.59      0.18      0.28      9720
       3       0.61      0.47      0.53     43571
       4       0.64      0.75      0.69     71630
       5       0.43      0.12      0.19     10103
       6       0.53      0.21      0.30      9578
       7       0.77      0.84      0.80    118829

accuracy                           0.75    385380
macro avg       0.60      0.44      0.47    385380
weighted avg       0.74      0.75      0.73    385380

XGBoost Model Results:

XGBoost Accuracy Score for Train Set: 0.678043562311387

XGBoost Accuracy Score for Test Set: 0.6795863822720432

XGBoost Classification Report:

           precision    recall  f1-score   support

       0       0.67      0.00      0.01       770
       1       0.76      0.88      0.82    121656
       2       0.63      0.05      0.10      9522
       3       0.52      0.29      0.37     43271
       4       0.64      0.61      0.62     70936
       5       0.27      0.00      0.00     10211
       6       0.43      0.05      0.08      9482
       7       0.64      0.82      0.72    119532

accuracy                           0.68    385380
macro avg       0.57      0.34      0.34    385380
weighted avg       0.65      0.68      0.64    385380

Neural Network Model Results:

Epoch 1/5 48173/48173 [==============================] - 123s 3ms/step - loss: 2.0535 - accuracy: 0.4762

Epoch 2/5 48173/48173 [==============================] - 122s 3ms/step - loss: 1.2062 - accuracy: 0.5730

Epoch 3/5 48173/48173 [==============================] - 123s 3ms/step - loss: 1.1845 - accuracy: 0.5816

Epoch 4/5 48173/48173 [==============================] - 122s 3ms/step - loss: 1.1672 - accuracy: 0.5899

Epoch 5/5 48173/48173 [==============================] - 123s 3ms/step - loss: 1.1525 - accuracy: 0.5959

Neural Network Accuracy Score for Test Set: 0.5619751933156885

Thus, it is clearly evident that KNN outperforms XGBoost and Dense Neural Network for this dataset.

Contact :

For any question, please contact

Lakshay Mehra: mehralakshay2@gmail.com