Crowd safety remains a critical challenge in public spaces where rapid population density changes can escalate into hazardous situations. Timely detection of risky behavior is therefore essential to enable early intervention and prevent escalation. To address this need, we present a real-time crowd risk detection framework that fuses motion dynamics with semantic scene interpretation to recognize hazards in surveillance scenarios involving dense crowds. The proposed system integrates two main modules: a Motion Entropy module, which computes entropy heatmaps of optical flow to capture anomalous crowd motions, and a Scene Semantics module, which leverages deep semantic segmentation to evaluate spatial risk based on visual context. The outputs of these modules are fused at both patch and frame levels, allowing the system to capture behavioral aberrations together with environmentally grounded risks. A Random Forest classifier is employed to assess risk levels, and alert thresholds are automatically determined using Otsu’s method. The model is evaluated on real-world vandalism surveillance footage and achieves an accuracy of 89.1%, precision of 87.5%, recall of 91.0%, and AUC of 0.90, outperforming individual feature-based baselines. The framework produces interpretable risk maps and real-time alerts, making it well-suited for deployment in open public spaces, event venues, and transportation hubs. These results demonstrate the effectiveness of multimodal fusion for situational awareness and early-warning systems in crowd safety applications.