DeepStream SOP — контроль регламентов на промышленном видео

★ 7.6 · content

deepstream-sop is a Claude Code skill that guides AI assistants in building, extending, and debugging the NVIDIA DeepStream SOP Inference Microservice — a GPU-accelerated FastAPI service (port 8300) that detects whether factory operators perform assembly-line steps in the correct order. The four-stage SOPVideoProcessor pipeline runs GEBD boundary detection (DDM model via Triton CAPI) in Stage 1, chunk segmentation in Stage 2, per-chunk VLM action classification via embedded vLLM (Cosmos Reason 1 or 2) in Stage 3, and SOP sequence validation in Stage 4. Input sources include video files, RTSP streams, and Basler/Pylon cameras via /v1/chat/completions; results are delivered through SSE streaming or Kafka NvProto/JSON output. Use this skill when working with SOPVideoProcessor internals, nvds_custom_postprocess, chunk-level latency measurement, TensorRT optimization, or Docker Compose deployment of the nvds-action-sop container.