Activity Recognition Algorithms from Accelerometer Signals (team project)

❖ Abstract

In this assignment, we were tasked with developing an activity recognition system from accelerometer signals using machine learning techniques. The process involves preprocessing the data by selecting a mobile phone position, converting the three-dimensional signal to one-dimensional, and handling sensor errors. Feature extraction is performed by dividing the signal into windows and extracting statistical measures and power spectrum estimates. Four classification models, k-nearest neighbors (KNN), SVM with RBF kernel, decision trees, and MLP neural networks, are employed. The evaluation follows the LOSO method, training the models on data from all but one participant and evaluating on the left-out participant’s data. The confusion matrix and accuracy are calculated to assess the classification performance. The report addresses questions regarding parameter optimization, confusion between activities, and the selection of the best classifier. Additionally, we collected our own accelerometer data and evaluated them using the models.

❖ Files Included

• Interactive Python Notebook - Notebook 1 (greek)
• Interactive Python Notebook - Notebook 2 (greek)
• Interactive Python Notebook - Notebook 3 (greek)
• Interactive Python Notebook - Notebook 4 (greek)
• The corresponding .csv files are included in the repository, in the respective folders.

❖ Keywords

Machine Learning, Activity Recognition, Accelerometer Signals, Preprocessing, Feature Extraction, SVM (Support Vector Machine), RBF (Radial Basis Function) Kernel, Decision Trees, K-Nearest Neighbors, MLP (Multilayer Perceptron) Neural Network, LOSO (Leave-One-Subject-Out) Evaluation, Confusion Matrix, Accuracy, Optimal Parameters, Classification Models, Performance Evaluation, Data Collection