Enhanced Activity Recognition Through Joint Utilization of Decimal Descriptors and Temporal Binary Motions

Abstract

Human Action Recognition (HAR) has been a prominent area of research within machine learning over the last few decades. Its applications span domains such as visual surveillance, robotics, and pedestrian detection. Despite the numerous techniques introduced by computer vision researchers to address HAR, persistent challenges include dealing with redundant features and computational cost. This paper specifically addresses the challenge of silhouette-based human activity recognition. While previous research on silhouette-based HAR has predominantly focused on recognition from a singular perspective, the aspect of view invariance has often been overlooked. This paper presents a novel framework that aims to achieve view-invariant Human Action Recognition. The proposed approach integrates a pre-processing stage based on the extraction of multiple 2D Differential History Binary Motions (DHBMs) from spatio-temporal frames capturing human motion. These multi-batch DHBMs are then used to capture and analyse human behaviour using the Decimal Descriptor Pattern (DDP) approach. This strategy enhances the extraction of intricate details from image data, contributing to a more robust HAR methodology. The selected features are processed by the Sparse Stacked Auto-encoder (SSAE), a representative of deep learning methods, to provide effective detection of human activity. The subsequent classification is performed using Softmax. The experiments are conducted on publicly available datasets, namely IXMAS and KTH. The results of the study demonstrate the superior performance of our methodology compared to previous approaches, achieving higher levels of accuracy.