DeepStream — генерация пайплайнов для видеоаналитики

★ 7.6 · content

deepstream-generate-pipeline is a Claude Code skill that builds ready-to-run gst-launch-1.0 pipelines for the NVIDIA DeepStream SDK through an interactive questionnaire covering seven parameters: input source, number of sources, inference type, tracker, sink, platform, and optional transforms. The retrieval backend uses BM25 with structural metadata boosting to score and rank results from 270+ verified pipelines, requiring only Python 3.8+ stdlib with zero external pip dependencies. Supported inference engines include nvinfer and Triton (nvinferserver); trackers cover NvDCF, IOU, NvSORT, and DeepSORT; inputs range from local MP4/H264/H265 files and images to RTSP streams and USB cameras; sinks include display, file save (MP4/JPG/PNG), RTSP out, and fakesink. Platforms supported are x86 dGPU (T4, A100, L40, RTX, etc.) and aarch64 Jetson (Orin, Xavier, Nano) / SBSA (Grace, GH200). A companion validate_pipeline.py script runs a four-stage check — syntax, elements, properties, and live parse — to verify the generated pipeline before deployment.