In response to these open challenges, this study aims to estimate a continuous Social Tension Index directly from raw video streams by integrating deep visual embeddings, short-range motion cues, and boosting-based regression within a computationally efficient framework. Through this approach, we achieve fine-grained monitoring of collective agitation and applicability to real-time evaluation on edge devices, as it provides an aggregated tension score rather than a binary label. The proposed design thus serves as an intermediate between traditional early warning and video-based macro analysis, supporting practical, proactive regional management and risk control \cite{bib24}.

To address these existing challenges, our research focuses on developing a continuous Social Tension Index that processes raw video data directly. We accomplish this by combining deep visual embeddings with short-range motion indicators and boosting-based regression techniques, all within a framework designed for computational efficiency. This method enables detailed tracking of group agitation levels and supports real-time assessment on edge computing devices, since it generates an overall tension score instead of simple binary classifications. Our proposed system therefore bridges the gap between conventional early warning systems and video-based macro-level analysis, enabling practical and proactive approaches to regional management and risk mitigation \cite{bib24}.