Skip to content

Nadim-Mahmud/Activity-detection-using-ML-model

BranchesTags

Repository files navigation

Activity Detection Machine Learning Project

Project Overview

This project focuses on detecting human activities using machine learning techniques. The dataset includes sensor data from 15 users performing various activities. The data is segmented and features are extracted using a sliding window approach. Multiple machine learning models, including logistic regression, decision trees, random forest, and neural networks, are trained and evaluated. The results indicate that Random Forest, XGBoost, and CNN models perform the best in recognizing activities.

Human Activity Recognition Dataset

This dataset contains activity tracking data collected from 15 users performing 15 different activities each. The activities include actions such as sitting, standing, walking, running, using a computer, and performing object manipulation tasks like picking up items from the floor. Each activity is labeled and categorized, with transition states included for specific scenarios (e.g., stationary to wearing a device). There are some missing data. The frequency of the sensors is 100Hz.

Dataset Details

The dataset includes data from 15 individuals (User IDs 13-27) performing various activities. These activities are categorized into sitting (reading, writing, typing, browsing, moving head/body, moving chair), standing (static standing, picking up items from the floor), dynamic (walking, running, taking stairs), and transitions (stationary to wearing a device, putting it back, and stationary with a 15-second delay). Each activity is documented with corresponding experiment numbers and labeled files. Some entries contain missing sensor data, such as accelerometer and gyroscope data.

Sensor Data Figures

To better understand the dataset, we provide visualizations of the sensor data collected from the accelerometer, gyroscope, and magnetometer. These figures are located in the figures folder.

Accelerometer Data

Shows raw acceleration forces in X, Y, and Z
axes during activities.

Gyroscope Data

Illustrates rotation rates around X, Y, and Z
axes.

Magnetometer Data

Presents magnetic field strength along X, Y, and Z
axes.

Feature Extraction Overview

This feature extraction process segments accelerometer and gyroscope data using a sliding window approach to compute statistical features for activity recognition. The data is segmented with configurable window sizes (e.g., 100, 200, 300, etc.) and overlap percentages (e.g., 25%, 50%). The step size is calculated as window_size * (1 - overlap). For each segment, the mean, median, variance, and standard deviation are computed for each axis (x, y, z) of the accelerometer and gyroscope data. The process involves segmenting the sensor data for each experiment ID, calculating the features for each window, and appending the user IDs. Missing data is handled appropriately, and the results are aggregated into a DataFrame. The feature files are saved as features/w{window_size}_o{overlap_percentage}_features.csv for multiple configurations.

Machine Learning Models and Techniques

The uploaded Jupyter Notebook indicates that various machine learning models were utilized. Here’s a list of models and techniques identified from the notebook:

Models Used:

  1. Logistic Regression - A linear classifier for baseline comparisons.
  2. Naive Bayes (GaussianNB) - Fast probabilistic model assuming feature independence.
  3. Support Vector Machines (SVC) - Effective for non-linear and high-dimensional data.
  4. Decision Tree Classifier - Interpretable tree-based model for classification.
  5. K-Nearest Neighbors (KNN) - Distance-based classifier for simple patterns.
  6. Random Forest Classifier - Ensemble method reducing overfitting.
  7. AdaBoost Classifier - Boosting weak learners for improved performance.
  8. Gradient Boosting Classifier - Sequentially reduces residual errors.
  9. XGBoost Classifier - High-performance gradient boosting implementation.
  10. Neural Networks (Keras):
    • Dense Layers - Fully connected layers for classification tasks.
    • CNN (Conv1D) - Captures spatial patterns in sequential data.
    • LSTM - Captures temporal dependencies in sequential data.
    • SimpleRNN - Handles simpler sequential patterns.

Supporting Techniques:

  • Standard Scaler - Normalizes data for consistent scale.
  • Principal Component Analysis (PCA) - Reduces dimensionality for efficiency.
  • Cross-Validation (KFold, StratifiedKFold) - Ensures robust model evaluation.
  • GridSearchCV and RandomizedSearchCV - Optimizes hyperparameters.
  • TSFresh - Automates feature extraction from time-series data.

Results

The results demonstrate that ensemble models such as Random Forest, XGBoost, and Gradient Boost consistently achieved the best performance, with accuracies reaching up to 68% and F1-scores of 0.68 in configurations with larger window sizes (e.g., 500) and higher overlaps (e.g., 50%). Among neural networks, Artificial Neural Networks (ANN) performed competitively, achieving up to 60% accuracy, but were generally outperformed by ensemble models. Larger window sizes and overlaps improved feature representation, enhancing model accuracy and robustness. Ensemble methods emerged as the most reliable classifiers for activity recognition, effectively capturing complex patterns in the data.

Window 100 & 25% Overlap

Model Accuracy Precision Recall F1 Score K-Fold Score Stratified Score
Logistic Regression 0.30 0.30 0.30 0.28 0.25 0.26
Decision Trees 0.51 0.51 0.51 0.51 0.601 0.603
Random Forest 0.64 0.64 0.64 0.64 0.72 0.72
Gaussian Naïve Bayes 0.22 0.23 0.22 0.19 0.13 0.13
Support Vector Classifier 0.56 0.55 0.56 0.55 0.275 0.274
KNN 0.52 0.54 0.52 0.51 0.36 0.365
AdaBoost 0.32 0.32 0.32 0.26 0.324 0.323
XGBoost 0.64 0.64 0.64 0.64 0.719 0.715
Gradient Boost 0.60 0.61 0.60 0.60 0.65 0.648
Artificial Neural Networks (ANN) 0.57 0.57 0.57 0.57 0.464 0.487

Window 100 & 50% Overlap

Model Accuracy Precision Recall F1 Score K-Fold Score Stratified Score
Logistic Regression 0.29 0.29 0.29 0.26 0.25 0.24
Decision Trees 0.53 0.54 0.53 0.54 0.62 0.63
Random Forest 0.68 0.69 0.68 0.68 0.76 0.77
Gaussian Naïve Bayes 0.26 0.26 0.26 0.22 0.13 0.13
Support Vector Classifier 0.58 0.58 0.58 0.57 0.29 0.29
KNN 0.55 0.56 0.55 0.53 0.37 0.37
AdaBoost 0.35 0.41 0.35 0.30 0.34 0.33
XGBoost 0.67 0.63 0.63 0.62 0.66 0.66
Gradient Boost 0.63 0.68 0.67 0.67 0.73 0.73
Artificial Neural Networks (ANN) 0.60 0.59 0.60 0.59 0.52 0.53

Window 200 & 25% Overlap

Model Accuracy Precision Recall F1 Score K-Fold Score Stratified Score
Logistic Regression 0.31 0.31 0.31 0.29 0.174 0.174
Decision Trees 0.63 0.63 0.63 0.62 0.72 0.72
Random Forest 0.47 0.48 0.47 0.47 0.591 0.593
Gaussian Naïve Bayes 0.25 0.26 0.25 0.22 0.26 0.259
Support Vector Classifier 0.55 0.55 0.55 0.55 0.25 0.255
KNN 0.49 0.53 0.49 0.48 0.349 0.349
AdaBoost 0.34 0.4 0.34 0.29 0.325 0.328
XGBoost 0.64 0.63 0.64 0.63 0.72 0.72
Gradient Boost 0.61 0.60 0.61 0.60 0.662 0.663
Artificial Neural Networks (ANN) 0.55 0.55 0.55 0.55 0.396 0.406

Window 200 & 50% Overlap

Model Accuracy Precision Recall F1 Score K-Fold Score Stratified Score
Logistic Regression 0.32 0.33 0.32 0.30 0.179 0.180
Decision Trees 0.54 0.53 0.54 0.53 0.631 0.633
Random Forest 0.68 0.68 0.68 0.67 0.766 0.768
Gaussian Naïve Bayes 0.24 0.27 0.24 0.21 0.153 0.152
Support Vector Classifier 0.59 0.59 0.59 0.58 0.277 0.277
KNN 0.53 0.55 0.53 0.51 0.369 0.371
AdaBoost 0.37 0.40 0.37 0.32 0.351 0.353
XGBoost 0.67 0.67 0.67 0.66 0.747 0.747
Gradient Boost 0.63 0.63 0.63 0.62 0.685 0.681
Artificial Neural Networks (ANN) 0.60 0.59 0.60 0.60 0.445 0.440

Window 300 & 25% Overlap

Model Accuracy Precision Recall F1 Score K-Fold Score Stratified Score
Logistic Regression 0.32 0.32 0.32 0.30 0.139 0.137
Decision Trees 0.47 0.46 0.47 0.46 0.585 0.580
Random Forest 0.62 0.62 0.62 0.61 0.724 0.726
Gaussian Naïve Bayes 0.30 0.31 0.30 0.27 0.149 0.151
Support Vector Classifier 0.54 0.54 0.54 0.53 0.236 0.236
KNN 0.48 0.52 0.48 0.46 0.345 0.344
AdaBoost 0.39 0.40 0.39 0.36 0.325 0.328
XGBoost 0.63 0.63 0.63 0.62 0.726 0.724
Gradient Boost 0.58 0.58 0.58 0.57 0.670 0.671
Artificial Neural Networks (ANN) 0.56 0.55 0.56 0.55 0.370 0.376

Window 300 & 50% Overlap

Model Accuracy Precision Recall F1 Score K-Fold Score Stratified Score
Logistic Regression 0.33 0.34 0.33 0.31 0.318 0.316
Decision Trees 0.50 0.50 0.50 0.50 0.635 0.637
Random Forest 0.66 0.66 0.66 0.66 0.771 0.770
Gaussian Naïve Bayes 0.28 0.31 0.28 0.25 0.275 0.275
Support Vector Classifier 0.58 0.57 0.58 0.57 0.256 0.255
KNN 0.51 0.54 0.51 0.49 0.372 0.372
AdaBoost 0.35 0.40 0.35 0.34 0.344 0.334
XGBoost 0.66 0.66 0.66 0.66 0.758 0.761
Gradient Boost 0.62 0.61 0.62 0.61 0.697 0.693
Artificial Neural Networks (ANN) 0.60 0.59 0.60 0.59 0.422 0.423

Window 400 & 25% Overlap

Model Accuracy Precision Recall F1 Score K-Fold Score Stratified Score
Logistic Regression 0.35 0.36 0.35 0.33 0.334 0.332
Decision Trees 0.46 0.46 0.46 0.46 0.582 0.580
Random Forest 0.61 0.61 0.61 0.60 0.725 0.723
Gaussian Naïve Bayes 0.30 0.33 0.30 0.28 0.289 0.289
Support Vector Classifier 0.54 0.54 0.54 0.53 0.222 0.223
KNN 0.47 0.51 0.47 0.45 0.349 0.351
AdaBoost 0.38 0.42 0.38 0.38 0.344 0.361
XGBoost 0.62 0.61 0.62 0.61 0.738 0.735
Gradient Boost 0.58 0.58 0.58 0.58 0.676 0.672
Artificial Neural Networks (ANN) 0.56 0.56 0.56 0.56 0.350 0.345

Window 400 & 50% Overlap

Model Accuracy Precision Recall F1 Score K-Fold Score Stratified Score
Logistic Regression 0.35 0.36 0.35 0.33 0.338 0.335
Decision Trees 0.51 0.51 0.51 0.51 0.628 0.630
Random Forest 0.66 0.65 0.66 0.65 0.776 0.775
Gaussian Naïve Bayes 0.31 0.34 0.31 0.28 0.291 0.290
Support Vector Classifier 0.59 0.59 0.59 0.59 0.253 0.252
KNN 0.49 0.53 0.49 0.48 0.376 0.379
AdaBoost 0.38 0.39 0.38 0.37 0.348 0.343
XGBoost 0.67 0.66 0.67 0.66 0.771 0.767
Gradient Boost 0.63 0.62 0.63 0.62 0.709 0.709
Artificial Neural Networks (ANN) 0.59 0.59 0.59 0.59 0.412 0.393

Window 500 & 25% Overlap

Model Accuracy Precision Recall F1 Score K-Fold Score Stratified Score
Logistic Regression 0.36 0.37 0.36 0.35 0.202 0.207
Decision Trees 0.46 0.46 0.46 0.45 0.590 0.589
Random Forest 0.60 0.60 0.60 0.60 0.732 0.733
Gaussian Naïve Bayes 0.29 0.32 0.29 0.27 0.172 0.173
Support Vector Classifier 0.55 0.55 0.55 0.54 0.218 0.214
KNN 0.45 0.50 0.45 0.43 0.353 0.356
AdaBoost 0.35 0.37 0.35 0.30 0.320 0.317
XGBoost 0.61 0.60 0.61 0.60 0.740 0.745
Gradient Boost 0.57 0.57 0.57 0.57 0.687 0.688
Artificial Neural Networks (ANN) 0.57 0.57 0.57 0.57 0.341 0.348

Window 500 & 50% Overlap

Model Accuracy Precision Recall F1 Score K-Fold Score Stratified Score
Logistic Regression 0.37 0.37 0.37 0.36 0.203 0.206
Decision Trees 0.51 0.51 0.51 0.51 0.630 0.630
Random Forest 0.66 0.66 0.66 0.66 0.780 0.780
Gaussian Naïve Bayes 0.31 0.34 0.31 0.29 0.310 0.308
Support Vector Classifier 0.58 0.58 0.58 0.57 0.241 0.240
KNN 0.49 0.53 0.49 0.48 0.373 0.378
AdaBoost 0.37 0.40 0.37 0.36 0.376 0.376
XGBoost 0.67 0.67 0.67 0.67 0.779 0.780
Gradient Boost 0.62 0.62 0.62 0.62 0.718 0.716
Artificial Neural Networks (ANN) 0.56 0.56 0.56 0.56 0.368 0.343

Deep Learning Results

For Window Size: 500, Overlap: 25%, the following results were achieved:

  • LSTM: 53.25% accuracy
  • CNN: 65.78% accuracy
  • RNN: 49.43% accuracy

Feature Extraction Using TSFresh

  • TSFresh extracted 2,351 features for each window configuration.
  • The number of rows generated:
    • Window 400, Overlap 25%: 19,146 rows
    • Window 400, Overlap 50%: 28,654 rows
    • Window 500, Overlap 25%: 15,287 rows
    • Window 500, Overlap 50%: 22,789 rows
  • PCA retained 95% variance with 5 features.

Conclusion

Random Forest, XGBoost, and CNN emerged as the most effective models, with CNN achieving the highest accuracy for sequential data.

About

Activity detection using ML final project

Resources

Readme
Activity

Stars

0 stars

Watchers

1 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

No packages published

Languages