Smart Detection Triage

Smart Detection Triage is a real-time Physical AI application running locally on the SiMa.ai Modalix DevKit. It continuously processes a camera or RTSP video stream using YOLO26m for object detection, then applies a configurable routing policy based on object class, confidence, and polygon zone. Instead of invoking an expensive vision-language model on every frame, the system selectively sends only important events—such as a high-confidence person detected in a restricted zone—to Qwen3-VL for structured scene enrichment. Vehicles can be suppressed, while unmatched or capacity-limited events are placed into a bounded review queue. The application produces schema-validated JSON alerts containing authoritative detector facts, the matched policy rule, zone, confidence, timestamps, latency measurements, and Qwen-generated visual context. Cooldown, deduplication, queue limits, asynchronous processing, and automatic stream recovery keep continuous detection responsive. A live dashboard allows operators to adjust the confidence threshold without restarting inference and observe FPS, trigger rate, avoided VLM requests, latency, alerts, and review activity. YOLO and Qwen are third-party model families deployed through SiMa-compatible Neat and LLiMa runtimes; the project’s key innovation is the policy-gated orchestration layer that determines when each model earns its computational cost.