A Deep Learning Framework for Single-Person Activity Recognition in Multi-Person Indoor Environments

Recognizing human activities in environments with multiple interacting individuals is crucial for various applications, including surveillance and human-computer interaction. Several challenges exist in human action recognition, such as activity detection in multiperson scenarios, view variation, and occlusion. This paper presents a solution for activity recognition in a multi-person environment. Due to limited accessibility and the availability of datasets in this scenario, a new dataset was recorded from 40 individuals (30 male, 10 female) in an indoor environment under good illumination conditions. Two different networks have been proposed: one for person detection and re-identification and another for activity classification. The activity classification employs a hybrid network utilizing image-based and human joint motion features. The proposed methodology was tested on state-of-the-art datasets, NTU-RGB+D, and MSR daily activity datasets. The model outperforms all state-of-the-art models, achieving an accuracy of 97.88%, 96.20%, and 96.70% for our generated, MSR activity and NTU RGB+D datasets, respectively.