Get Started (Standalone Mode / pymavlink)#

This guide provides a step-by-step walkthrough for testing the UAV Vision Analytics application in standalone mode (pymavlink) and running the demo with a simulated UAV camera feed/RealSense cameras.

How It Works#

A self-contained stack. PX4 SITL, MAVLink router, MQTT broker, and Metrics Manager are all started together (docker-compose-pymavlink.yml). Telemetry flows from PX4 SITL through mavlink-router to the DL Streamer container, where pymavlink reads it directly over UDP.

uav vision analytics standalone

Telemetry flow:

        sequenceDiagram
    participant PX4 as PX4 SITL
    participant RTR as mavlink-router
    participant OVL as gvapython (MavlinkReceiver)
    participant Frame as Video Frame

    PX4->>RTR: MAVLink stream (UDP :14550)
    RTR->>OVL: broadcast UDP :14541
    Note over OVL: background thread parses<br/>GLOBAL_POSITION_INT, VFR_HUD,<br/>GPS_RAW_INT into latest_data
    Frame->>OVL: process_frame() per frame
    OVL->>Frame: ROI labels (ALT · SPD · HDG · LAT · LON · SATS)

Services:

Service	Image	Ports	Role
`dlstreamer-pipeline-server`	`intel/dlstreamer-pipeline-server` + pymavlink	`8081`, `8555`	AI inference, RTSP output
`px4`	`px4io/px4-sitl`	`14550`	Flight controller simulator
`mavlink-router`	custom build	`14551`	MAVLink UDP routing (:14550 → :14541)
`metrics-manager`	`intel/metrics-manager`	`9090`	CPU/GPU/NPU/power metrics

Steps to Test the Application#

System Requirements#

See System Requirements for the full list of software and hardware prerequisites.

1. Configure environment#

Clone the repo and Get into the directory:

git clone https://github.com/open-edge-platform/edge-ai-suites.git --branch release-2026.2.0
cd edge-ai-suites/federal-and-aerospace-ai-suite/uav-vision-analytics

make init

make init creates .env from the template and auto-detects your Intel GPU device paths (GPU_DEVICE, GPU_RENDER_DEVICE), Intel NPU (NPU_DEVICE), and Intel RealSense / USB camera (REALSENSE_DEVICE). It skips if .env already exists.

Then set your host IP address in .env:

nano .env   # set HOST_IP=<your-machine-IP>

2. Prepare the model#

Download and export the YOLOv8n-VisDrone model to OpenVINO FP16 IR:

make model

See the AI Model guide for model details.

3. Standalone mode (pymavlink)#

make pymav-up

4. Start inference pipelines#

Two options are available depending on your use case:

Option A — Managed RTSP output (recommended)#

Runs pipeline_manager.py inside the DLSPS container. It monitors the drone’s ARMED/DISARMED state and automatically starts and stops inference pipelines. Annotated frames are served as RTSP on port 8555.

make start-rtsp starts one device pipeline at a time (default: GPU). Pass DEVICE=cpu|gpu|npu|all to choose:

make start-rtsp                # GPU only (default)
make start-rtsp DEVICE=cpu     # CPU only
make start-rtsp DEVICE=npu     # NPU only
make start-rtsp DEVICE=all     # CPU + GPU + NPU simultaneously

Note: Open QGroundControl (QGC) to connect and press takeoff, which arms the UAV (Only arming will automatically disarm the UAV after a few seconds). The pipeline manager will automatically start the selected pipeline and serve annotated RTSP streams.

DEVICE=npu requires NPU_DEVICE to have been detected during make init — falls back to GPU otherwise.

Refer to the QGroundControl guide for instructions on connecting to the RTSP stream.

pymavlink mode — output streams (only the selected DEVICE is active, unless DEVICE=all):

rtsp://<HOST_IP>:8555/uav-mavlink-cpu    (CPU pipeline)
rtsp://<HOST_IP>:8555/uav-mavlink-gpu    (GPU pipeline)
rtsp://<HOST_IP>:8555/uav-mavlink-npu    (NPU pipeline) # If NPU Device is available

File-source pipelines (started via REST API or benchmark script) — output path is set in the POST request body (e.g. uav-mavlink-cpu for the uav_object_detection_cpu pipeline).

Option B — Manual REST API#

Start a single pipeline directly without the pipeline manager. Useful for testing individual pipelines or custom configurations.

# CPU pipeline
INSTANCE_ID=$(curl -s -X POST \
  http://localhost:8081/pipelines/user_defined_pipelines/uav_object_detection_cpu \
  -H "Content-Type: application/json" \
  -d '{
    "destination": {
      "metadata": {
        "type": "file",
        "path": "/tmp/results.jsonl",
        "format": "json-lines"
      },
      "frame": {
        "type": "rtsp",
        "path": "uav-mavlink-cpu"
      }
    },
    "parameters": {
      "detection-properties": {
        "model": "/home/pipeline-server/resources/models/yolov8n-visdrone/best_openvino_model/best.xml",
        "device": "CPU"
      }
    }
  }' | tr -d '"')
echo "Instance ID: $INSTANCE_ID"

Change following three values to switch between CPU / GPU / NPU:

Pipeline name in the URL path (uav_object_detection_cpu → _gpu / _npu)
RTSP path in the request body (uav-mavlink-cpu → uav-mavlink-gpu / uav-mavlink-npu)
Device in detection-properties (CPU → GPU / NPU)

View the annotated stream immediately after posting:

ffplay rtsp://<HOST_IP>:8555/uav-mavlink-cpu   # or uav-mavlink-gpu / uav-mavlink-npu

Stop a pipeline:

curl -X DELETE http://localhost:8081/pipelines/${INSTANCE_ID}

5. View the output stream#

# Install ffmpeg if not present, then view any device stream
ffplay rtsp://<HOST_IP>:8555/uav-mavlink-cpu   # CPU
ffplay rtsp://<HOST_IP>:8555/uav-mavlink-gpu   # GPU
ffplay rtsp://<HOST_IP>:8555/uav-mavlink-npu   # NPU

The annotated stream includes bounding boxes for detected objects (person, car, bus, truck, van, bicycle, tricycle, awning-tricycle, motor, others) and a live telemetry overlay (GPS, altitude, speed, heading).

6. Stop all services#

Stop and remove the standalone pymavlink stack (also removes named volumes):

make pymav-down

Pipelines#

pymavlink mode (`config-pymavlink.json`)#

Pipeline	Device	Source	Output
`uav_object_detection_cpu`	CPU	Looped video file (`uav_sample.avi`)	RTSP `:8555`
`uav_object_detection_gpu`	GPU	Looped video file (`uav_sample.avi`)	RTSP `:8555`
`uav_object_detection_npu`	NPU	Looped video file (`uav_sample.avi`)	RTSP `:8555`
`uav_realsense_cpu`	CPU	Intel RealSense camera (v4l2src)	RTSP `:8555`
`uav_realsense_gpu`	GPU	Intel RealSense camera (v4l2src)	RTSP `:8555`
`uav_realsense_npu`	NPU	Intel RealSense camera (v4l2src)	RTSP `:8555`

Telemetry Overlay Fields#

Each output frame carries these overlaid fields in the upper-left corner:

Field	Source MAVLink message	Description
`Name`	—	Name passed as argument to the gvapython
`Frame`	—	Running frame counter
`ALT`	`GLOBAL_POSITION_INT.relative_alt`	Relative altitude (m)
`SPD`	`VFR_HUD.groundspeed`	Ground speed (m/s)
`HDG`	`GLOBAL_POSITION_INT.hdg`	Heading (degrees)
`LAT`	`GPS_RAW_INT.lat`	Latitude
`LON`	`GPS_RAW_INT.lon`	Longitude
`SATS`	`GPS_RAW_INT.satellites_visible`	GPS satellites visible

Port Reference#

Port	Protocol	Service	Mode
`8081`	HTTP	DL Streamer REST API	All modes
`8555`	RTSP	Annotated video output	All modes
`14541`	UDP	MAVLink broadcast (mavlink-router)	pymavlink modes
`9090`	HTTP	metrics-manager (HW metrics)	pymavlink modes

RealSense Camera Support#

Intel RealSense camera setup and pipelines details are provided in the RealSense guide.

Documentation#

Document	Description
index.md	Application overview and component block diagrams
benchmark.md	Performance benchmarking guide (`calc_stream_density.sh`)
makefile.md	Makefile target reference
troubleshooting.md	Known issues and resolutions