Video analytics, configured in YAML.
Vision SDK turns a single config file into a production video pipeline — inputs, models, tracking, and outputs — running on NVIDIA DeepStream.
Inputs flow through models to outputs.
Every pipeline has the same shape. Point it at any stream, wire in any combination of detection, tracking, and classification, and send the result anywhere.
One file. One runtime. Six places to send the result.
Production primitives, ready to compose.
YAML configuration
Describe streams, models, tracking, and outputs in one file. No Python glue, no DeepStream plumbing.
Model chaining
Compose detectors, trackers, and classifiers in any order. Classifiers operate on parent detections by reference.
Multi-stream outputs
Fan cameras into tiled outputs, or split them into per-stream files, RTSP, or RTMP endpoints.
Real-time analytics
Per-frame detection JSON to Kafka — ready for dashboards, alerting, and downstream services.
One config. One command.
Install the wheel inside a DeepStream 8.0 container, write a YAML file, and run it. That's the whole thing.
pipeline:
name: "My First Pipeline"
streams:
- uri: "rtsp://192.168.1.100:554/live"
models:
- name: "detector"
type: "detector"
architecture: "yolo"
onnx_file: /path/to/model.onnx
label_file: /path/to/labels.txt
sinks:
width: 1280
height: 720
outputs:
- type: "display"
$ vision-runner pipeline.yaml