Decoding the Digital Gaze: GuardianView and the Algorithmic Analysis of Children's Visual Consumption

Abstract

The rapid proliferation of digital streaming platforms has intensified concerns regarding children's exposure to inappropriate content, necessitating robust, automated safeguarding solutions. This paper introduces GuardianView, a novel, proactive, and privacy-preserving framework engineered for real-time domestic violence detection on edge devices. We propose a Triple-Path Inference Architecture that harmonizes computational efficiency with semantic depth. This pipeline integrates: (i) a Normalized Cross-Correlation (NCC) heuristic gatekeeper for rapid, low-power static signature filtering; (ii) a YOLO-based deep neural network for spatial feature extraction; and (iii) a Temporal Action Recognition (TAR) module to capture complex, context-dependent behavioral dynamics.
Experimental evaluations conducted in a macOS environment demonstrate that GuardianView achieves a Mean Average Precision (mAP) of 92%, outperforming baseline detection models by 7% while maintaining a deterministic latency of < 30 ms per frame. By implementing an intelligent "Early Exit" logic, the framework significantly minimizes computational overhead, restricting CPU utilization to < 5% during background monitoring. Furthermore, we address the ethical imperatives of on-device intelligence by ensuring that all telemetry and inference remain strictly localized, thereby mitigating the privacy risks inherent in cloud-based surveillance. This research provides a scalable, ethically aligned roadmap for next-generation parental control systems, establishing both a novel technical architecture and a design paradigm for future human-centric AI safety systems.