Depth-Aware, Asynchronous Video Analytics for Production: The Intellivision Platform

ABSTRACT IntelliVision is a production-ready platform for video analytics and face authentication supporting people counting, vehicle counting and automatic number plate recognition (ANPR), parking analysis, emergency flow detection, lobby/crowd thresholding, pothole detection, wildlife/pest monitoring, food-waste estimation, and room readiness analysis. We present a generalizable methodology - asynchronous orchestration with job queues, GPU-aware inference, model selection with resilient fallbacks, and task-specific post-processing - aimed at robustness under constrained resources. The methodology incorporates (i) depth-aware detector selection that adapts to scene structure; (ii) lifetime filtering plus hierarchical identity merging to reduce ID switches and overcounting in crowded scenes; and (iii) slot-gated admission and lazy model initialization to manage resource constraints. We discuss design trade-offs and limitations. Implementation details (APIs, deployment, developer tooling) are intentionally minimized to foreground the research contributions. Keywords: video analytics, object detection, tracking, ANPR, depth estimation, Celery, GPU, YOLO, RT-DETR, Re-ID, Qdrant, InsightFace