Codebook-Based Feature Engineering for Human Activity Recognition Using Multimodal Sensory Data

Recently, Human Activity Recognition (HAR) using sensory data from various devices has become increasingly vital in fields like healthcare, elderly care, and smart home systems. However, many existing HAR systems face challenges such as high computational demands or the need for large datasets. This paper introduces a codebook-based approach designed to overcome these challenges by offering a more efficient method for HAR with reduced computational costs. Initially, the raw time series data is segmented intosmaller subsequences, and codebooks are constructed using the Bag of Features (BOF) approach. Each subsequence is then assigned softly based on the center of each cluster (codeword), resulting in a histogram-based feature vector. These encoded feature vectors are subsequently classified using a Support Vector Machine (SVM). The proposed method was evaluated using the OPPORTUNITY dataset, comprising data from 72 sensors, achieving a classification accuracy of 90.7%. In comparison to other advanced techniques, our approach not only demonstrated superior accuracy in recognizing human activities but also significantly reduced computational costs. The use of soft assignments for mapping codewords to subsequences efficiently captured the key patterns within the activity data. The findings validate that the proposed codebook-based method provides substantial improvements in both accuracy and efficiency for HAR systems.