Human action-recognition systems enhance public safety by detecting abnormal behavior autonomously. RGB sensors commonly used in such systems capture personal information of subjects and, as a result, run the risk of potential privacy leakage. On the other hand, privacy-safe alternatives, such as depth or thermal sensors exhibit poorer performance because they lack the semantic context provided by RGB sensors. Moreover, the data availability of privacy-safe alternatives is significantly lower than RGB sensors. To address these problems, we explore effective cross-modality distillation methods in this paper, aiming to distill the knowledge of context-rich large-scale pre-trained RGB-based models into privacy-safe depth-based models. Based on extensive experiments on multiple architectures and benchmark datasets, We propose an effective method for training privacy-safe depth-based action-recognition models via cross-modality distillation: cross-modality mixing distillation. This approach improves both the performance and efficiency by enabling interaction between depth and RGB modalities through a linear combination of their features. By utilizing the proposed cross-modal mixing distillation approach, we achieve state-of-the-art accuracy in two depth-based action recognition benchmarks. The code and the pre-trained models will be available upon publication.

Subway systems are highly vulnerable to disasters because of their confined underground structures and limited evacuation routes, making accurate evacuation analysis essential for reducing casualties. Most existing studies overlook panic behavior, leading to unrealistic assessments of evacuation efficiency. This study develops a modeling framework that integrates panic behavior into agent-based subway evacuation simulations. The framework incorporates three behavioral factors: reaction delays to emergency cues, hesitation at decision points, and irrational route choices. Simulation experiments were conducted under different occupancy conditions, and results from panic-integrated and non-panic scenarios were compared. Findings show that the inclusion of panic significantly prolongs evacuation time, with delays of up to 30% in full-scale scenarios due to congestion and route errors. These outcomes demonstrate that panic behavior exerts a decisive influence on evacuation dynamics and should not be neglected in simulation studies. Incorporating panic into evacuation modeling provides a more realistic basis for designing safer subway systems and developing effective emergency response strategies.

This study presents a novel system for accurate camera-to-person distance estimation in CCTV environments. To address the limitations of existing approaches—which often require extensive training data and lack object-level precision—we propose a hybrid framework that integrates SAM’s zero-shot segmentation with monocular depth estimation. Our method isolates human subjects from complex backgrounds and incorporates Kernel Density Estimation (KDE), log-space learning, and linear residual blocks to improve prediction accuracy. This approach is designed to resolve the non-linear mapping between visual features and metric distances. Evaluations on a custom dataset demonstrate a mean absolute error (MAE) of 0.65 m on 1612 test images, using only 30 training samples. Notably, the use of SAM for fine-grained segmentation significantly outperforms conventional bounding box methods, reducing the MAE from 0.82 m to 0.65 m. The proposed system offers immediate applicability to security surveillance and disaster response scenarios, with its minimal data requirements enhancing its practical deployability.