Real-time face recognition systems are crucial in enhancing situational awareness, particularly in public safety scenarios where accurate individual identification can significantly influence critical decisions. Such environments are inherently unpredictable, with face images captured under non-ideal conditions involving varying angles, lighting, and motion. This paper presents a robust face recognition pipeline for public safety applications. The proposed pipeline combines YOLO-based face detection, FaceNet-based feature extraction, and high-performance vector retrieval using Qdrant. This work implements an embedding accumulation strategy that delays identity confirmation until recognition evidence is aggregated over several frames, reducing false positives while prioritizing precision over throughput. We evaluated the model’s performance across camera-to-subject distances, an important yet often overlooked factor in real-world deployments. We also evaluated the pipeline with and without the accumulation strategy using a hybrid dataset of CelebA, LFW, and custom-recorded video scenarios featuring controlled head movements and varying distances. The accumulation strategy extends the two-meter effective recognition range while preserving high precision. The results demonstrate the pipeline’s potential to significantly improve public safety situational awareness.