Skeleton-based feature selection and activity recognition

Abstract

This thesis investigates how Skeleton-based Human Activity Recognition (HAR) can be improved by utilizing cross-joint interactions to increase activity classification performance. Skeleton-based HAR efficiently characterizes body motion while preserving privacy. The SCOUT sequential pattern-mining tool, based on the SPAM algorithm, is used to capture cross-landmark patterns spanning multiple skeletal joints. To mitigate the high computational expense of sequential mining, various analyses for extracting meaningful skeletal features are conducted: Statistical Co-variance Analysis, Random Forest (RF) feature selection, and Graph Convolutional Networks (GCNs). The resulting filtered joints are then utilized in SCOUT. Additionally, several activity identification systems are investigated, including a hybrid RF + LSTM model, a late fusion model, and GCN-based system. Overall performance in recognizing Activities of Daily Living (ADLs) is compared against baseline SCOUT. Experimental findings reveal that deep learning models yield better classification performance than traditional pattern mining on complex ADLs.