deepstream-api-jetson

DeepStream + FastAPI backend for Jetson Orin NX.

Features

• RTSP camera ingestion
• NVIDIA PeopleNet detection
• Multi-object tracking
• Line crossing analytics
• Zone dwell time analytics
• JSON metadata for backend integration

Target: Jetson Orin NX (JetPack 6.x)

Repository Structure

deepstream-api-jetson/
│
├── README.md
├── docker/
│   ├── Dockerfile
│   └── docker-compose.yml
│
├── app/
│   ├── main.py
│   ├── pipeline.py
│   ├── model_manager.py
│   └── config/
│       ├── deepstream_app.txt
│       ├── pgie_peoplenet.txt
│       ├── tracker_config.yml
│       └── analytics_config.txt
│
├── models/
│   └── peoplenet/
│
└── scripts/
    ├── jetson_check.sh
    └── build_engine.sh

---

1. Hardware Requirements

• Jetson Orin NX (16GB recommended)
• NVMe SSD recommended
• 25W or MAXN power mode

---

2. Fresh Jetson Setup

Flash using NVIDIA SDK Manager:

• JetPack 6.x
• Ubuntu 22.04

Verify installation:

cat /etc/nv_tegra_release

---

3. Set Performance Mode

sudo nvpmodel -m 0
sudo jetson_clocks

---

4. Validate Software Stack

bash scripts/jetson_check.sh

---

5. Build and Run Container

docker compose build
docker compose up

---

6. Start Pipeline

POST request:

POST /start
Body: { "rtsp_url": "rtsp://camera_ip/stream" }

---

7. Model Management

Place models in:

models/peoplenet/

Build TensorRT engine on device:

bash scripts/build_engine.sh

Always build engines on the target Jetson device.

---

8. Performance Target (Orin NX 25W)

  Streams   Expected FPS
  --------- -----------------
  1         30+ FPS
  4         30 FPS each
  8         15--20 FPS each

Use FP16 for stable deployment.