Get Started#

The Image-Based Video Search is a sample application that demonstrates how developers can leverage edge AI technologies to solve real-world challenges. It enables efficient processing and searching of video data to identify objects of interest, providing actionable insights in real-time.

This application showcases searching through video data generated by cameras around a city to find the closest match to the user-provided image.

By following this guide, you will learn how to:

  • Set up the sample application: Use Docker Compose to quickly deploy the application in your environment.

  • Run a predefined pipeline: Execute a sample pipeline to see real-time traffic monitoring and object detection in action.

  • Modify application parameters: Customize settings like input sources and detection thresholds to adapt the application to your specific requirements.

Prerequisites#

Set up and First Use#

  1. Clone the Repository and Build Containers:

    • Create and navigate to directory:

        git clone https://github.com/open-edge-platform/edge-ai-suites.git
  cd edge-ai-suites/metro-ai-suite/image-based-video-search
  docker compose build

  1. Download the Models:

    • Download the models

      Linux Instructions 
      cd src/dlstreamer-pipeline-server

# Define a generic directory for storing models
MODELS_PATH="$(pwd)/models"

# Create the directory if it doesn't exist
mkdir -p "$MODELS_PATH"

# Download and convert the ResNet-50 model
docker run --rm \
    --user=root \
    -e http_proxy -e https_proxy -e no_proxy \
    -v "$MODELS_PATH:/output" \
    openvino/ubuntu22_dev:2024.6.0 bash -c \
    "omz_downloader --name resnet-50-pytorch --output_dir models && \
    omz_converter --name resnet-50-pytorch --download_dir models --output_dir models && \
    cp -r ./models/public/resnet-50-pytorch /output && \
    chown -R $(id -u):$(id -g) /output"

# Download and convert the Person-Vehicle-Bike Detection model
docker run --rm \
    --user=root \
    -e http_proxy -e https_proxy -e no_proxy \
    -v "$MODELS_PATH:/output" \
    openvino/ubuntu22_dev:2024.6.0 bash -c \
    "omz_downloader --name person-vehicle-bike-detection-2004 --output_dir models && \
    omz_converter --name person-vehicle-bike-detection-2004 --download_dir models --output_dir models && \
    cp -r ./models/intel/person-vehicle-bike-detection-2004 /output && \
    chown -R $(id -u):$(id -g) /output"
      Windows Instructions 
      cd src/dlstreamer-pipeline-server

docker pull openvino/ubuntu22_dev:2024.6.0
$MODELS_PATH="$PWD\models"

docker run --rm `
    -e http_proxy -e https_proxy -e no_proxy \
    -v ${MODELS_PATH}:/output `
    openvino/ubuntu22_dev:2024.6.0 bash -c `
    "omz_downloader --name resnet-50-pytorch --output_dir models && `
    omz_converter --name resnet-50-pytorch --download_dir models --output_dir models && `
    cp -r ./models/public/resnet-50-pytorch /output"

docker run --rm `
    -e http_proxy -e https_proxy -e no_proxy \
    -v ${MODELS_PATH}:/output `
    openvino/ubuntu22_dev:2024.6.0 bash -c `
    "omz_downloader --name person-vehicle-bike-detection-2004 --output_dir models && `
    omz_converter --name person-vehicle-bike-detection-2004 --download_dir models --output_dir models && `
    cp -r ./models/intel/person-vehicle-bike-detection-2004 /output"
      ```

</details>

  2. Update DOCKER_REGISTRY variable in .env file Note: This step is required to deploy with the pre-release images

    • The recommended setting to use pre-release images is: DOCKER_REGISTRY=ghcr.io/open-edge-platform/edge-ai-libraries/ Please remember to include / at the end.

  3. Start the Application:

    • Go back to the folder of compose.yml and run the application using Docker Compose:

      cd ../..
docker compose up -d

  4. Verify the Application:

    • Check that the application is running:

      docker compose ps

  5. Access the Application:

    • Open a browser and go to the following endpoints to access the application:

      • Stream UI: http://localhost:8889/stream

      • App UI: http://localhost:3000

      • Search UI: http://localhost:9000/docs

      • MilvusDB UI: http://localhost:8000/

  6. Run the Application:

    • Analyze Stream: Use the predefined video and click Analyze Stream to start processing the video stream.

    • Video Search: Click the Upload Image button to upload your own images for search or click the Capture Frame button to capture and adjust frames from the video stream. Click the Search Object button.

    • Expected Results:

      • Matched search results, including metadata, timestamps, distance to show the confidence rate of the prediction, and frames that include detected objects (e.g., vehicles, pedestrians, bikes).

    image1

    image2

Make Changes#

  1. Change Object Detection and Object Classification Models

    • To use your own models instead of the default models, follow these steps:

      • Open the config.json file.

      • Change the paths in the pipeline section to point to your own models. Replace the paths for gvadetect and gvaclassify with the paths to your models:

        {
    "config": {
        "logging": {
            "C_LOG_LEVEL": "INFO",
            "PY_LOG_LEVEL": "INFO"
        },
        "pipelines": [
            {
                "name": "filter-pipeline",
                "source": "gstreamer",
                "queue_maxsize": 50,
                "pipeline": "{auto_source} name=source ! decodebin ! video/x-raw ! videoconvert ! gvadetect model=/models/your-detection-model/FP32/your-detection-model.xml model-proc=/models/your-detection-model/your-detection-model.json inference-interval=3 threshold=0.4 model-instance-id=detect1 device=CPU ! queue ! gvainference model=/models/your-classification-model/FP32/your-classification-model.xml inference-region=1 name=classification model-instance-id=infer1 device=CPU ! queue ! videoconvertscale ! gvametaconvert add-tensor-data=true name=metaconvert ! jpegenc ! appsink name=destination",
                "auto_start": false,
                "mqtt_publisher": {
                    "publish_frame": true,
                    "topic": "edge_video_analytics_results"
                }
            },
            {
                "name": "search_image",
                "source": "image_ingestor",
                "queue_maxsize": 50,
                "pipeline": "appsrc name=source  ! decodebin ! videoconvert ! gvainference model=/models/your-classification-model/FP32/your-classification-model.xml model-instance-id=infer2 device=CPU ! gvametaconvert add-tensor-data=true ! appsink name=destination"
            }
        ]
    }
}

  2. Change Input Video:

    • To use your own video instead of the default sample video, follow these steps:

      • Open the compose.yml file and locate the streaming-pipeline service.

      • Change the URL in the command section to point to your own video file. Replace https://github.com/intel-iot-devkit/sample-videos/raw/master/person-bicycle-car-detection.mp4 with the URL of your video file:

        services:
  streaming-pipeline:
    ...
    command: >
      bash -c "
        wget -O file.mp4 <your-video-url> && \
        gst-launch-1.0 filesrc location=file.mp4 ! qtdemux name=mdemux ! h264parse ! video/x-h264,stream-format=byte-stream ! mpegtsmux name=mux ! filesink location=file.ts && \
        gst-launch-1.0 multifilesrc location=file.ts loop=true ! tsdemux ! h264parse ! rtspclientsink protocols=tcp location=rtsp://rtsp-server:8554/stream
      "
    ...

  3. Adjust Feature Matching Confidence:

    • Open the compose.yml file and locate the feature-matching service.

    • Change the CONFIDENCE_THRESHOLD to adjust the confidence level for feature matching:

      services:
  feature-matching:
    ...
    environment:
      ...
      CONFIDENCE_THRESHOLD: 0.7
    ...

    • Effect: Increasing the CONFIDENCE_THRESHOLD will make the feature matching more stringent, reducing false positives but potentially missing some true positives. Decreasing it will make the matching more lenient, increasing the chances of detecting true positives but also increasing false positives.

  4. Save Changes and Restart:

    • Save the file and restart the application:

      docker compose restart

  5. Verify Updates:

    • Expected Results:

      • The application processes data from the updated input source.

      • Detection results align with the changed models

    • Confirm changes through:

      • Logs:

        docker compose logs

Troubleshooting#

  1. Containers Not Starting:

    • Check the Docker logs for errors:

      docker compose logs

  2. Port Conflicts:

    • Update the ports section in the Compose file to resolve conflicts.

  3. ibvs-milvusdb container is unhealthy:

    • Currently, milvusdb does not work with proxy servers. Make sure that the proxies http_proxy, https_proxy and no_proxy are set to empty string in compose.yml file

  4. Empty search results after clicking on Search Object:

    • Make sure the models are able to detect the objects in the stream correctly

    • Make sure you have analysed the stream first to capture the video frames into milvus database

    • Make sure you are using the right frame to search the object

    • Increase the ‘To’ timestamp in the search results to accommodate the latest results

Supporting Resources#