Automatic criminal threat detection via body-worn cameras using deep learning technique: A systematic literature review

This systematic literature review (SLR) examines the application of deep learning within boy-worn camera (BWC) systems for enhancing real-time threat detection and situational awareness in law enforcement and public safety. The review covers peer-reviewed studies published between 2019 and 2025, following PRISMA guidelines. A total of 53 studies from Scopus, Web of Science, IEEE Xplore, and SpringerLink were included in the analysis. Finding reveals a strong reliance on convolutional neural network (CNNs) and lightweight architecture like YOLO and MobileNet for object detection and violence recognition. While these models perform well in controlled environment, their deployment in mobile surveillance faces challenges such as motion distortion, occlusion, and hardware limitations. Emotion recognition and behavioral prediction are also explored, but model accuracy remains sensitive to poor quality and partial visibility. A significant limitation identified is the scarcity of BWC-specific datasets. The FALEB multi-modal dataset represents a notable effort to provide domain-specific data, supporting task like facial recognition, action detection, and gaze analysis within BWC contexts. Computational constraints further hinder real-time implementation, especially on embedded devices. The review underscores the urgent need for more BWC-focused datasets, comprehensive evaluation protocols, and privacy-aware models designs. While deep learning offers great promise for BWC systems, addressing these practical, ethical barriers is essential for effective real-world applications.