Outlier Detection in Human Activity Recognition Systems

The paper focuses on the detection of outliers in human motion phases. The aim was to find the most effective machine learning method to detect anomalous segments within physical activities. The article investigates the effectiveness of machine learning algorithms in detecting outlier activities within datasets. A novel approach employing nested binary classifier models is proposed to enhance outlier detection. The models’ nested binary classifiers were evaluated for accuracy and precision in identifying outlier activities and compared with classifiers k-nearest neighbor, support vector machine, CART decision trees, and naive Bayes classifier. The nested models, iteratively refined through multiple nesting levels, demonstrate improved accuracy compared to standalone classifiers, particularly in identifying outlier activities. Results indicate varying performance across datasets and nesting levels, highlighting the potential of nested models in enhancing anomaly detection in diverse applications.