Performance Evaluation of Open-Set Classification in Human Activity Recognition via a Residual Neural Network Architecture

A novel technique using residual neural networks is introduced in this research to address the open-set classification problem in the field of Human Activity Classification (HAR). The open-set issue presents several problems when performing classification tasks due to the removal or lack of labels during the model training phase. Limited amounts of research has been conducted to address this issue directly despite its recurrence in real-world applications. The Time Series Classification Residual Network (TSCResNet) deep learning model introduced in this study is evaluated with the benchmark PAMAP2 Physical Activity Monitoring Dataset. By using the same quantitative methods and data preprocessing protocols as previous research in the field of open-set HAR, results show that the TSCResNet model architecture is able achieve improved classification results.