Share via

Quickstart: Track objects in a live video

  • Cikk

edge icon
Alternatively, check out topics under Create video applications in the service.

Note

Azure Video Analyzer has been retired and is no longer available.

Azure Video Analyzer for Media is not affected by this retirement. It is now rebranded to Azure Video Indexer. Click here to read more.

This quickstart shows you how to use Azure Video Analyzer edge module to track objects in a live video feed from a (simulated) IP camera. You will see how to apply a computer vision model to detect objects in a subset of the frames in the live video feed. You can then use an object tracker node to track those objects in the other frames.

The object tracker comes in handy when you need to detect objects in every frame, but the edge device does not have the necessary compute power to be able to apply the vision model on every frame. If the live video feed is at, say 30 frames per second, and you can only run your computer vision model on every 15th frame, the object tracker takes the results from one such frame, and then uses optical flow techniques to generate results for the 2nd, 3rd,…, 14th frame, until the model is applied again on the next frame.

This quickstart uses an Azure VM as an IoT Edge device, and it uses a simulated live video stream.

Prerequisites

  • An Azure account that includes an active subscription. Create an account for free if you don't already have one.

    Note

    You will need an Azure subscription where you have access to both Contributor role, and User Access Administrator role. If you do not have the right permissions, please reach out to your account administrator to grant you those permissions.

  • Visual Studio Code, with the following extensions:

    Tip

    When you're installing the Azure IoT Tools extension, you might be prompted to install Docker. Feel free to ignore the prompt.

  • .NET Core 3.1 SDK.

Set up Azure resources

Deploy to Azure

The deployment process will take about 20 minutes. Upon completion, you will have certain Azure resources deployed in the Azure subscription, including:

  1. Video Analyzer account - This cloud service is used to register the Video Analyzer edge module, and for playing back recorded video and video analytics.
  2. Storage account - For storing recorded video and video analytics.
  3. Managed Identity - This is the user assigned managed identity used to manage access to the above storage account.
  4. Virtual machine - This is a virtual machine that will serve as your simulated edge device.
  5. IoT Hub - This acts as a central message hub for bi-directional communication between your IoT application, IoT Edge modules and the devices it manages.

In addition to the resources mentioned above, following items are also created in the 'deployment-output' file share in your storage account, for use in quickstarts and tutorials:

  • appsettings.json - This file contains the device connection string and other properties needed to run the sample application in Visual Studio Code.
  • env.txt - This file contains the environment variables that you will need to generate deployment manifests using Visual Studio Code.
  • deployment.json - This is the deployment manifest used by the template to deploy edge modules to the simulated edge device.

Tip

If you run into issues creating all of the required Azure resources, please use the manual steps in this quickstart.

Overview

Track objects in live video

This diagram shows how the signals flow in this quickstart. An edge module simulates an IP camera hosting a Real-Time Streaming Protocol (RTSP) server. An RTSP source node pulls the video feed from this server and sends video frames to the HTTP extension processor node.

The HTTP extension node plays the role of a proxy. It converts every 15th video frame to the specified image type. Then it relays the image over HTTP to another edge module that runs an AI model behind an HTTP endpoint. In this example, that edge module uses the YOLOv3 model which can detect many types of objects. The HTTP extension processor node receives the detection results and sends these results and all the video frames (not just the 15th frame) to the object tracker node. The object tracker node uses optical flow techniques to track the object in the 14 frames that did not have the AI model applied to them. The tracker node publishes its results to the IoT Hub message sink node. This IoT Hub message sink node then sends those events to IoT Edge Hub.

Note

You should review the discussion of the about the trade-off between accuracy and processing power with the object tracker node.

In this quickstart, you will:

  1. Setup your development environment.
  2. Deploy the required edge modules.
  3. Create and deploy the live pipeline.
  4. Interpret the results.
  5. Clean up resources.

Set up your development environment

Get the sample code

  1. Clone the AVA C# samples repository.

  2. Start Visual Studio Code, and open the folder where the repo has been downloaded.

  3. In Visual Studio Code, browse to the src/cloud-to-device-console-app folder and create a file named appsettings.json. This file contains the settings needed to run the program.

  4. Browse to the file share in the storage account created in the setup step above, and locate the appsettings.json file under the "deployment-output" file share. Click on the file, and then hit the "Download" button. The contents should open in a new browser tab, which should look like:

    {
    "IoThubConnectionString" : "HostName=xxx.azure-devices.net;SharedAccessKeyName=iothubowner;SharedAccessKey=XXX",
    "deviceId" : "avasample-iot-edge-device",
    "moduleId" : "avaedge"
}

    The IoT Hub connection string lets you use Visual Studio Code to send commands to the edge modules via Azure IoT Hub. Copy the above JSON into the src/cloud-to-device-console-app/appsettings.json file.

  5. Next, browse to the src/edge folder and create a file named .env. This file contains properties that Visual Studio Code uses to deploy modules to an edge device.

  6. Browse to the file share in the storage account created in the setup step above, and locate the env.txt file under the "deployment-output" file share. Click on the file, and then hit the "Download" button. The contents should open in a new browser tab, which should look like:

         SUBSCRIPTION_ID="<Subscription ID>"
     RESOURCE_GROUP="<Resource Group>"
     AVA_PROVISIONING_TOKEN="<Provisioning token>"
     VIDEO_INPUT_FOLDER_ON_DEVICE="/home/localedgeuser/samples/input"
     VIDEO_OUTPUT_FOLDER_ON_DEVICE="/var/media"
     APPDATA_FOLDER_ON_DEVICE="/var/lib/videoanalyzer"
     CONTAINER_REGISTRY_USERNAME_myacr="<your container registry username>"
     CONTAINER_REGISTRY_PASSWORD_myacr="<your container registry password>"

    Copy the JSON from your env.txt into the src/edge/.env file.

Connect to the IoT Hub

  1. In Visual Studio Code, set the IoT Hub connection string by selecting the More actions icon next to the AZURE IOT HUB pane in the lower-left corner. Copy the string from the src/cloud-to-device-console-app/appsettings.json file.

    Note

    You might be asked to provide Built-in endpoint information for the IoT Hub. To get that information, in Azure portal, navigate to your IoT Hub and look for Built-in endpoints option in the left navigation pane. Click there and look for the Event Hub-compatible endpoint under Event Hub compatible endpoint section. Copy and use the text in the box. The endpoint will look something like this: Endpoint=sb://iothub-ns-xxx.servicebus.windows.net/;SharedAccessKeyName=iothubowner;SharedAccessKey=XXX;EntityPath=<IoT Hub name>

  2. In about 30 seconds, refresh Azure IoT Hub in the lower-left section. You should see the edge device avasample-iot-edge-device, which should have the following modules deployed:

    • Edge Hub (module name edgeHub)
    • Edge Agent (module name edgeAgent)
    • Video Analyzer (module name avaedge)
    • RTSP simulator (module name rtspsim)

Prepare to monitor the modules

When you use run this quickstart or tutorial, events will be sent to the IoT Hub. To see these events, follow these steps:

  1. Open the Explorer pane in Visual Studio Code, and look for Azure IoT Hub in the lower-left corner.

  2. Expand the Devices node.

  3. Right-click on avasample-iot-edge-device, and select Start Monitoring Built-in Event Endpoint.

    Note

    You might be asked to provide Built-in endpoint information for the IoT Hub. To get that information, in Azure portal, navigate to your IoT Hub and look for Built-in endpoints option in the left navigation pane. Click there and look for the Event Hub-compatible endpoint under Event Hub compatible endpoint section. Copy and use the text in the box. The endpoint will look something like this: Endpoint=sb://iothub-ns-xxx.servicebus.windows.net/;SharedAccessKeyName=iothubowner;SharedAccessKey=XXX;EntityPath=<IoT Hub name>

Deploy the required modules

  1. In Visual Studio Code, right-click the src/edge/deployment.yolov3.template.json file and then select Generate IoT Edge Deployment Manifest.

    Generate IoT Edge Deployment Manifest

  2. The deployment.yolov3.amd64.json manifest file is created in the src/edge/config folder.

  3. Right-click src/edge/config/deployment.yolov3.amd64.json and select Create Deployment for Single Device.

    Create Deployment for Single Device

  4. When you're prompted to select an IoT Hub device, select avasample-iot-edge-device.

  5. After about 30 seconds, in the lower-left corner of the window, refresh Azure IoT Hub. The edge device now shows the following deployed modules:

    • The Video Analyzer edge module, named avaedge.

    • The rtspsim module, which simulates an RTSP server and acts as the source of a live video feed.

    • The yolov3 module, which uses the YOLOV3 model to detect a variety of objects

      List of deployed IoT Edge modules

Create and deploy the live pipeline

Examine and edit the sample files

In Visual Studio Code, browse to the src/cloud-to-device-console-app folder. Here you'll see the appsettings.json file that you created along with a few other files:

  • c2d-console-app.csproj: The project file for Visual Studio Code.
  • operations.json: This file lists the different operations that you would run.
  • Program.cs: The sample program code, which:
    • Loads the app settings.
    • Invokes direct methods exposed by Video Analyzer module. You can use the module to analyze live video streams by invoking its direct methods.
    • Pauses for you to examine the output from the program in the TERMINAL window and the events generated by the module in the OUTPUT window.
    • Invokes direct methods to clean up resources.

  1. Edit the operations.json file:

    • Change the link to the pipeline topology:
    • "pipelineTopologyUrl" : "https://raw.githubusercontent.com/Azure/video-analyzer/main/pipelines/live/topologies/object-tracking/topology.json"
    • Under livePipelineSet, edit the name of the pipeline topology to match the value in the preceding link:
    • "topologyName" : "ObjectTrackingWithHttpExtension"
    • Under pipelineTopologyDelete, edit the name:
    • "name" : "ObjectTrackingWithHttpExtension"

Open the URL for the pipeline topology in a browser, and examine the settings for the HTTP extension node.

   "samplingOptions":{
       "skipSamplesWithoutAnnotation":"false",
       "maximumSamplesPerSecond":"2"
   }

Here, skipSamplesWithoutAnnotation is set to false because the extension node needs to pass through all frames, whether or not they have inference results, to the downstream object tracker node. The object tracker is capable of tracking objects over 15 frames, approximately. Your AI model has a maximum FPS for processing, which is the highest value that maximumSamplesPerSecond should be set to.

Run the sample program

  1. To start a debugging session, select the F5 key. You see messages printed in the TERMINAL window.

  2. The operations.json code starts off with calls to the direct methods pipelineTopologyList and livePipelineList. If you cleaned up resources after you completed previous quickstarts, then this process will return empty lists and then pause. To continue, select the Enter key.

    -------------------------------Executing operation pipelineTopologyList-----------------------  
Request: pipelineTopologyList  --------------------------------------------------
{
"@apiVersion": "1.1"
}
---------------  
Response: pipelineTopologyList - Status: 200  ---------------
{
"value": []
}
--------------------------------------------------------------------------
Executing operation WaitForInput

Press Enter to continue

    The TERMINAL window shows the next set of direct method calls:

    • A call to pipelineTopologySet that uses the contents of pipelineTopologyUrl
    • A call to livePipelineSet that uses the following body:
    {
  "@apiVersion": "1.1",
  "name": "Sample-Pipeline-1",
  "properties": {
    "topologyName": "ObjectTrackingWithHttpExtension",
    "description": "Sample pipeline description",
    "parameters": [
      {
        "name": "rtspUrl",
        "value": "rtsp://rtspsim:554/media/camera-300s.mkv"
      },
      {
        "name": "rtspUserName",
        "value": "testuser"
      },
      {
        "name": "rtspPassword",
        "value": "testpassword"
      }
    ]
  }
}
    • A call to livePipelineActivate that activates the live pipeline and the flow of video.
    • A second call to livePipelineList that shows that the live pipeline is in the running state.

  3. The output in the TERMINAL window pauses at a Press Enter to continue prompt. Do not press Enter yet. Scroll up to see the JSON response payloads for the direct methods you invoked.

  4. Switch to the OUTPUT window in Visual Studio Code. You see messages that the Video Analyzer edge module is sending to the IoT hub. The following section of this quickstart discusses these messages.

  5. The live pipeline continues to run and print results. The RTSP simulator keeps looping the source video. To stop the live pipeline, return to the TERMINAL window and select Enter.

  6. The next series of calls cleans up resources:

    • A call to livePipelineDeactivate deactivates the live pipeline.
    • A call to livePipelineDelete deletes the live pipeline.
    • A call to pipelineTopologyDelete deletes the pipeline topology.
    • A final call to pipelineTopologyList shows that the list is empty.

Interpret results

When you run the live pipeline, the results from the HTTP extension processor node pass through the IoT Hub message sink node to the IoT hub. The messages you see in the OUTPUT window contain a body section and an applicationProperties section. For more information, see Create and read IoT Hub messages.

In the following messages, the Video Analyzer module defines the application properties and the content of the body.

MediaSessionEstablished event

When a live pipeline is activated, the RTSP source node attempts to connect to the RTSP server that runs on the rtspsim-live555 container. If the connection succeeds, then the following event is printed. The event type is MediaSessionEstablished.

[IoTHubMonitor] [9:42:18 AM] Message received from [avasample-iot-edge-device/avaedge]:
{  "body": {
    "sdp": "SDP:\nv=0\r\no=- 1586450538111534 1 IN IP4 nnn.nn.0.6\r\ns=Matroska video+audio+(optional)subtitles, streamed by the LIVE555 Media Server\r\ni=media/camera-300s.mkv\r\nt=0 0\r\na=tool:LIVE555 Streaming Media v2020.03.06\r\na=type:broadcast\r\na=control:*\r\na=range:npt=0-300.000\r\na=x-qt-text-nam:Matroska video+audio+(optional)subtitles, streamed by the LIVE555 Media Server\r\na=x-qt-text-inf:media/camera-300s.mkv\r\nm=video 0 RTP/AVP 96\r\nc=IN IP4 0.0.0.0\r\nb=AS:500\r\na=rtpmap:96 H264/90000\r\na=fmtp:96 packetization-mode=1;profile-level-id=4D0029;sprop-parameter-sets=Z00AKeKQCgC3YC3AQEBpB4kRUA==,aO48gA==\r\na=control:track1\r\n"
  },
  "applicationProperties": {
    "dataVersion": "1.0",
    "topic": "/subscriptions/{subscriptionID}/resourceGroups/{name}/providers/microsoft.media/videoAnalyzers/{ava-account-name}",
    "subject": "/edgeModules/avaedge/livePipelines/Sample-Pipeline-1/sources/rtspSource",
    "eventType": "Microsoft.VideoAnalyzer.Diagnostics.MediaSessionEstablished",
    "eventTime": "2020-04-09T16:42:18.1280000Z"
  }
}

In this message, notice these details:

  • The message is a diagnostics event. MediaSessionEstablished indicates that the RTSP source node (the subject) connected with the RTSP simulator and has begun to receive a (simulated) live feed.
  • In applicationProperties, subject indicates that the message was generated from the RTSP source node in the live pipeline.
  • In applicationProperties, eventType indicates that this event is a diagnostics event.
  • The eventTime indicates the time when the event occurred.
  • The body contains data about the diagnostics event. In this case, the data comprises the Session Description Protocol (SDP) details.

Object tracking events

The HTTP extension processor node sends the 0th, 15th, 30th, … etc. frames to the yolov3 module, and receives the inference results. It then sends these results and all video frames to the object tracker node. Suppose an object was detected on frame 0 – then the object tracker will assign a unique sequenceId to that object. Then, in frames 1, 2,…,14, if it can track that object, it will output a result with the same sequenceId. In the following snippets from the results, note how the sequenceId is repeated, but the location of the bounding box has changed, as the object is moving.

From frame M:

  {
    "type": "entity",
    "subtype": "objectDetection",
    "inferenceId": "4d325fc4dc7a43b2a781bf7d6bdb3ff0",
    "sequenceId": "0999a1dde5b241c3a0b2db025f87ab32",
    "entity": {
      "tag": {
        "value": "car",
        "confidence": 0.95237225
      },
      "box": {
        "l": 0.0025893003,
        "t": 0.550063,
        "w": 0.1086607,
        "h": 0.12116724
      }
    }
  },

From frame N:

{
  "type": "entity",
  "subtype": "objectDetection",
  "inferenceId": "317aafdab7e940388be1e4c4cc58c366",
  "sequenceId": "0999a1dde5b241c3a0b2db025f87ab32",
  "entity": {
    "tag": {
      "value": "car",
      "confidence": 0.95237225
    },
    "box": {
      "l": 0.0027777778,
      "t": 0.54901963,
      "w": 0.108333334,
      "h": 0.12009804
    }
  }
},

Clean up resources

If you want to try other quickstarts or tutorials, keep the resources that you created. Otherwise, go to the Azure portal, go to your resource groups, select the resource group where you ran this quickstart, and delete all the resources.

Next steps