SMART-vision: survey of modern action recognition techniques in vision

2024-12-21

Multimedia tools and applications

AI Summary

To address key challenges in human activity recognition (HAR)—including opaque hybrid methodologies, unclear deep learning mechanisms, and poor generalization to unseen activities—this paper systematically reviews over 120 state-of-the-art works published between 2018 and 2024. We propose the SMART taxonomy—Semantic, Multi-scale, Adaptive, Robust, and Temporal—as the first unified classification framework to standardize evaluation and elucidate technological evolution. Focusing on emerging paradigms such as Transformers, spatiotemporal graph neural networks, multimodal fusion, self-supervised contrastive learning, and prompt-based fine-tuning, we rigorously analyze their impacts on classification accuracy, model interpretability, and lightweight deployment. We identify three persistent bottlenecks: data bias, limitations in temporal modeling, and insufficient cross-domain generalization. Finally, we articulate design principles for scalable, benchmark-driven evaluation, offering both theoretical foundations and practical guidelines to advance HAR research and deployment.

Technology Category

• Planning, Routing, and Scheduling: Activity and Plan Recognition
• Machine Learning: Hardware-aware ML
• Computer Vision: Video Understanding & Activity Analysis

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web

Problem

• Human Activity Recognition
• Deep Learning Integration
• Unknown Action Classification

Innovation

• SMART-Vision
• Deep Learning
• Open-HAR