A Hybrid ART-RBF Network Architecture for Viewpoint Invariant Human Activity Recognition

There is growing interest in the problem of human activity recognition, motivated by its countless promising applications in many domains. Despite much progress, most researchers have narrowed the problem towards fixed camera viewpoint owing to inherent difficulty to train their systems across all possible viewpoints. Fixed viewpoint systems are impractical in real scenarios. Therefore, we attempt to relax the infamous fixed viewpoint assumption and present a novel, efficient and biologically-inspired framework to recognize and classify human activities from monocular video source from arbitrary viewpoint. The proposed framework comprises two stages: human pose recognition and human activity recognition. We cascade an ensemble of invariant pose models and activity models hierarchically. All the models operate simultaneously in parallel and perform inference on impinging patterns that come from lower level. Pose models operate in a hybrid 3-layered bottom-up neural architecture. Activity models employ fuzzy-state hidden Markov model to classify activities. We have built a small-scale architecture for a proof-of-concept and performed some experiments on two publicly available datasets. The satisfactory experimental results demonstrate the efficacy of our framework and encourage us to further develop a full-scale architecture.