Human motion recognition (HMR) is an important application area in smart surveillance, medical surveillance, and human-computer interfaces. However, human activities are extremely diverse, and the spatiotemporal modeling is a computationally resource-demanding concept; thus, proper and effective recognition is difficult. Traditional deep learning networks are usually not generated automatically; they are computationally expensive or cannot be used to model a wide variety of motion patterns. The current paper introduces a Progressive Neural Architecture Search (PNAS) neural architecture, PNAS-HMR, that is an automatic human movement recognition system that finds the most effective deep network models without compromising recognition accuracy, computation cost, or interpretability. The given framework comprises a progressive search strategy, where the lightweight network candidates are taken into account, and an upward search strategy is used to construct more complex structures with the help of multi-objective search. The search is directed by a surrogate predictor of performance based on a search using Pareto-optimal architectures that meet accuracy, latency, and performance energy requirements. Assessment was performed using motion data from standard datasets, including NTU RGB + D 120, Kinetics-400, and UCF101. The architectures found were further refined and studied using visualization tools and ablation experiments. Experimental tests show that PNAS-HMR achieves higher recognition accuracy (up to 4%) with lower FLOPs and latency than state-of-the-art methods, by more than 30%. The architectures searched exhibit strong generalization and cross-dataset transferability. PNAS-HMR is an effective combination of neural architecture search and spatiotemporal learning that provides an energy-efficient, interpretable, and scalable framework for real-time human motion recognition in various practical problems.
Wang et al. (Tue,) studied this question.