A Shape-Based Feature Descriptor for Multi-View Human Activity Recognition

In recent decades, Human Activity Recognition (HAR) has become an eminent research topic due to its application in healthcare, surveillance, sports activity monitoring, etc. Despite the generous methods that have been created for HAR, recognizing human activity remains an arduous task due to noise, inter-class similarity, intra-class variability, overfitting, and high dimensionality. In this paper, we have proposed a shape-based feature descriptor, which is based on shape information and spatial structure of an image with a dimensionality reduction technique has been proposed. Further, this reduced feature is fed into the Support Vector Machine (SVM) classifier to perform activity recognition. The model has been evaluated on KTH, WEIZMANN, and IXMAS datasets for demonstrating superior performance against the current approaches. It has achieved an accuracy of $97.48 \%, 100 \%$, and 100% on the datasets, respectively. Through this work, we contribute powerful Machine Learning (ML)-based solutions that address overfitting and high dimensionality problems of HAR.