Fusion of Vision Based Features for Human Activity Recognition

Abstract

The field of image recognition has made significant progress in recent years, particularly in the domain of Human Activity Recognition (HAR). HAR systems use computer vision and machine learning algorithms to classify images into different activities. A review of existing literature in this area was conducted, focusing on a vision-based system for recognition purposes. The present study focuses on recognizing Eating, Walking, and Chitchat activities using a vision-based approach. The dataset consisted of 1306 images, equally divided between the three activities, with around 400-500 per activity. This work has potential applications in surveillance, assisted living, elder care, healthcare monitoring systems, human-robot interactions, and gaming and entertainment. The model can also be embedded for monitoring elderly persons who are living alone and also to keep the track of their activities. The best possible results were obtained by using the Random Forest classifier after using the Fusion of Binary Robust Invariant Scalable Keypoints (BRISK) and Scale Invariant Feature Transform (SIFT) feature detector, with an overall accuracy of 89.41% and the precision of 89.42%, if the fusion of these features is not taken into consideration then solely Random Forest classifier provided the highest accuracy of 86.98% when applied on BRISK extracted features. The main finding of the study is fusion of BRISK and SIFT feature descriptors which showed promising results for all classifiers and they act as a novel approach in this proposed model and research.