Deploy with Helm#

Prerequisites#

  • System Requirements

  • Kubernetes Cluster: Ensure you have a properly installed and configured Kubernetes cluster.

  • Tools Installed: Install the required tools:

    • Kubernetes CLI (kubectl)

    • Helm 3 or later

  • For Helm installation, refer to Helm website

  • Intel NFD and Device Plugins (required for GPU/NPU workloads): Install Node Feature Discovery (NFD) and the Intel GPU/NPU device plugins to enable hardware detection and scheduling. This ensures pods requesting GPU or NPU resources are only deployed on nodes with available hardware. Refer to release tags for available versions (tested with v0.35.0):

    # Pick a release version compatible with your cluster
export RELEASE_VERSION=v0.35.0

# Step 1: Create namespace for the Intel device plugins
kubectl create namespace intel-device-plugins

# Step 2: Allow privileged pods in the device plugin namespace
# Required because the plugin needs hostPath mounts and access to host device files.
kubectl label namespace intel-device-plugins \
  pod-security.kubernetes.io/enforce=privileged \
  pod-security.kubernetes.io/audit=privileged \
  pod-security.kubernetes.io/warn=privileged \
  --overwrite

# Step 3: Install Node Feature Discovery (NFD)
# NFD uses its own namespace: node-feature-discovery
kubectl apply -k "https://github.com/intel/intel-device-plugins-for-kubernetes/deployments/nfd?ref=${RELEASE_VERSION}"

# Step 4: Allow privileged pods in the NFD namespace
kubectl label namespace node-feature-discovery \
  pod-security.kubernetes.io/enforce=privileged \
  pod-security.kubernetes.io/audit=privileged \
  pod-security.kubernetes.io/warn=privileged \
  --overwrite

# Step 5: Install Intel GPU NodeFeatureRules
# These rules let NFD detect and label Intel GPU nodes.
kubectl apply -k "https://github.com/intel/intel-device-plugins-for-kubernetes/deployments/nfd/overlays/node-feature-rules?ref=${RELEASE_VERSION}"

# Step 6: Verify NFD pods are running
kubectl get pods -n node-feature-discovery

# Step 7: Verify the node got Intel GPU and NPU labels
kubectl get node $(hostname) --show-labels | tr ',' '\n' | grep intel

# Step 8: Install the Intel GPU device plugin
kubectl apply -n intel-device-plugins -k "https://github.com/intel/intel-device-plugins-for-kubernetes/deployments/gpu_plugin/overlays/nfd_labeled_nodes?ref=${RELEASE_VERSION}"

# Step 9: Install the Intel NPU device plugin
kubectl apply -n intel-device-plugins -k "https://github.com/intel/intel-device-plugins-for-kubernetes/deployments/npu_plugin/overlays/nfd_labeled_nodes?ref=${RELEASE_VERSION}"

    Verify the Intel Device Plugin pods are running:

    kubectl get pods -n intel-device-plugins

    Verify the GPU and NPU resources are advertised on nodes:

    kubectl get nodes -o json | jq '.items[] | {name: .metadata.name, gpu: .status.allocatable["gpu.intel.com/i915"], npu: .status.allocatable["npu.intel.com/accel"]}'

    Note: If your node uses Intel Xe discrete GPUs (Arc), set gpu: to .status.allocatable["gpu.intel.com/xe"].

Setup the application#

Note: The following instructions assume Kubernetes is already running in the host system with Helm package manager installed.

  1. Clone the edge-ai-suites repository and change into industrial-edge-insights-vision directory. The directory contains the utility scripts required in the instructions that follows.

    git clone https://github.com/open-edge-platform/edge-ai-suites.git -b release-2026.1.0
cd edge-ai-suites/manufacturing-ai-suite/industrial-edge-insights-vision/

  2. Set app specific values.yaml file.

    cp helm/values_worker-safety-gear-detection.yaml helm/values.yaml

  3. Optional: Pull the Helm chart and replace the existing Helm folder with it

    Note: The Helm chart should be downloaded when you are not using the helm chart provided in edge-ai-suites/manufacturing-ai-suite/industrial-edge-insights-vision/helm

    • Download helm chart with the following command

      helm pull oci://registry-1.docker.io/intel/worker-safety-gear-detection --version 1.3.0

    • unzip the package using the following command

      tar -xvf worker-safety-gear-detection-1.3.0.tgz

    • Replace the helm directory

      rm -rf helm && mv worker-safety-gear-detection helm

  4. Edit the HOST_IP, proxy and other environment variables in helm/values.yaml as follows

    env:
    HOST_IP: <HOST_IP>   # host IP address
    MINIO_ACCESS_KEY: <DATABASE USERNAME> #  example: minioadmin
    MINIO_SECRET_KEY: <DATABASE PASSWORD> #  example: minioadmin
    http_proxy: <http proxy> # proxy details if behind proxy
    https_proxy: <https proxy>
    SAMPLE_APP: worker-safety-gear-detection # application directory
webrtcturnserver:
    username: <username>  # WebRTC credentials e.g. intel1234
    password: <password>

    Note: To run the pipeline on GPU, set gpu.enabled:true in values.yaml. To run the pipeline on NPU, set npu.enabled:true - this also requires a GPU resource since NPU pipelines use VA-API (GPU) for video decoding. For Intel Arc (Xe) discrete GPUs, set gpu.type: "gpu.intel.com/xe".

  5. Install prerequisites. Run with sudo if needed.

    ./setup.sh helm

    This sets up application prerequisites, download artifacts, sets executable permissions for scripts, etc. Downloaded resource directories.

Deploy the application#

  1. Install the Helm chart

    helm install app-deploy helm -n apps --create-namespace

    After installation, check the status of the running pods:

    kubectl get pods -n apps

    To view logs of a specific pod, replace <pod_name> with the actual pod name from the output above:

    kubectl logs -n apps -f <pod_name>

  2. Copy the resources such as video and model from local directory to the dlstreamer-pipeline-server pod to make them available for application while launching pipelines.

    # Below is an example for Worker safety gear detection. Please adjust the source path of models and videos appropriately for other sample applications.

POD_NAME=$(kubectl get pods -n apps -o jsonpath='{.items[*].metadata.name}' | tr ' ' '\n' | grep deployment-dlstreamer-pipeline-server | head -n 1)

kubectl cp resources/worker-safety-gear-detection/videos/Safety_Full_Hat_and_Vest.avi $POD_NAME:/home/pipeline-server/resources/videos/ -c dlstreamer-pipeline-server -n apps

kubectl cp resources/worker-safety-gear-detection/models/* $POD_NAME:/home/pipeline-server/resources/models/ -c dlstreamer-pipeline-server -n apps

  3. Fetch the list of pipeline loaded available to launch

    ./sample_list.sh helm

    This lists the pipeline loaded in DL Streamer Pipeline Server.

    Output:

    # Example output for Worker Safety gear detection
Environment variables loaded from [WORKDIR]/manufacturing-ai-suite/industrial-edge-insights-vision/.env
Running sample app: worker-safety-gear-detection
Checking status of dlstreamer-pipeline-server...
Server reachable. HTTP Status Code: 200
Loaded pipelines:
[
    ...
    {
        "description": "DL Streamer Pipeline Server pipeline",
        "name": "user_defined_pipelines",
        "parameters": {
        "properties": {
            "detection-properties": {
            "element": {
                "format": "element-properties",
                "name": "detection"
            }
            }
        },
        "type": "object"
        },
        "type": "GStreamer",
        "version": "worker_safety_gear_detection"
    }
    ...
]

  4. Start the sample application with a pipeline.

    ./sample_start.sh helm -p worker_safety_gear_detection

    This command looks for the payload for the pipeline specified in -p argument above, inside the payload.json file and launches a pipeline instance in DL Streamer Pipeline Server. Refer to the table, to learn about the different options available.

    Output:

    # Example output for Worker Safety gear detection
Environment variables loaded from [WORKDIR]/manufacturing-ai-suite/industrial-edge-insights-vision/.env
Running sample app: worker-safety-gear-detection
Checking status of dlstreamer-pipeline-server...
Server reachable. HTTP Status Code: 200
Loading payload from [WORKDIR]/manufacturing-ai-suite/industrial-edge-insights-vision/helm/apps/worker-safety-gear-detection/payload.json
Payload loaded successfully.
Starting pipeline: worker_safety_gear_detection
Launching pipeline: worker_safety_gear_detection
Extracting payload for pipeline: worker_safety_gear_detection
Found 1 payload(s) for pipeline: worker_safety_gear_detection
Payload for pipeline 'worker_safety_gear_detection' {"source":{"uri":"file:///home/pipeline-server/resources/videos/Safety_Full_Hat_and_Vest.avi","type":"uri"},"destination":{"frame":{"type":"webrtc","peer-id":"worker_safety"}},"parameters":{"detection-properties":{"model":"/home/pipeline-server/resources/models/worker-safety-gear-detection/deployment/Detection/model/model.xml","device":"CPU"}}}
Posting payload to REST server at http://<HOST_IP>:30107/pipelines/user_defined_pipelines/worker_safety_gear_detection
Payload for pipeline 'worker_safety_gear_detection' posted successfully. Response: "74bebe7a5d1211f08ab0da88aa49c01e"

    Note: This starts the pipeline. You can view the inference stream on WebRTC by opening a browser and navigating to https://<HOST_IP>:30443/mediamtx/worker_safety/ for Worker Safety gear detection. If you are running Helm using an NGINX_HTTPS_PORT other than the default 30443, replace 30443 with <NGINX_HTTPS_PORT>.

Starting GPU and NPU based pipelines#

For GPU and NPU based pipelines, ensure you have done the necessary setup from here, and start the respective pipelines as following.

For GPU-based pipelines:

./sample_start.sh helm -p worker_safety_gear_detection_gpu

For NPU-based pipelines:

./sample_start.sh helm -p worker_safety_gear_detection_npu

  1. Get status of pipeline instance(s) running.

    ./sample_status.sh helm

    This command lists status of pipeline instances launched during the lifetime of sample application.

    Output:

    # Example output for Worker Safety gear detection
Environment variables loaded from [WORKDIR]/manufacturing-ai-suite/industrial-edge-insights-vision/.env
Running sample app: worker-safety-gear-detection
[
{
    "avg_fps": 30.036955894826452,
    "elapsed_time": 3.096184492111206,
    "id": "784b87b45d1511f08ab0da88aa49c01e",
    "message": "",
    "start_time": 1752100724.3075056,
    "state": "RUNNING"
}
]

  2. Stop pipeline instance.

    ./sample_stop.sh helm

    This command will stop all instances that are currently in RUNNING state and respond with the last status.

    Output:

    # Example output for Worker Safety gear detection
No pipelines specified. Stopping all pipeline instances
Environment variables loaded from [WORKDIR]/manufacturing-ai-suite/industrial-edge-insights-vision/.env
Running sample app: worker-safety-gear-detection
Checking status of dlstreamer-pipeline-server...
Server reachable. HTTP Status Code: 200
Instance list fetched successfully. HTTP Status Code: 200
Found 1 running pipeline instances.
Stopping pipeline instance with ID: 784b87b45d1511f08ab0da88aa49c01e
Pipeline instance with ID '784b87b45d1511f08ab0da88aa49c01e' stopped successfully. Response: {
    "avg_fps": 29.985911953641363,
    "elapsed_time": 37.45091152191162,
    "id": "784b87b45d1511f08ab0da88aa49c01e",
    "message": "",
    "start_time": 1752100724.3075056,
    "state": "RUNNING"
}

    If you wish to stop a specific instance, you can provide it with an --id argument to the command. For example, ./sample_stop.sh helm --id 784b87b45d1511f08ab0da88aa49c01e

  3. Uninstall the helm chart.

    helm uninstall app-deploy -n apps

Storing frames to S3 storage#

Applications can take advantage of the S3 publish feature from DL Streamer Pipeline Server and use it to save frames to an S3 compatible storage.

  1. Run all the steps mentioned in the section above to setup the application.

  2. Install the helm chart.

    helm install app-deploy helm -n apps --create-namespace

  3. Copy the resources such as video and model from local directory to the dlstreamer-pipeline-server pod to make them available for application while launching pipelines.

    # Below is an example for Worker safety gear detection. Please adjust the source path of models and videos appropriately for other sample applications.

POD_NAME=$(kubectl get pods -n apps -o jsonpath='{.items[*].metadata.name}' | tr ' ' '\n' | grep deployment-dlstreamer-pipeline-server | head -n 1)

kubectl cp resources/worker-safety-gear-detection/videos/Safety_Full_Hat_and_Vest.avi $POD_NAME:/home/pipeline-server/resources/videos/ -c dlstreamer-pipeline-server -n apps

kubectl cp resources/worker-safety-gear-detection/models/* $POD_NAME:/home/pipeline-server/resources/models/ -c dlstreamer-pipeline-server -n apps

  4. Install the package boto3 in your python environment if not installed.

    It is recommended to create a virtual environment and install it there. You can run the following commands to add the necessary dependencies as well as create and activate the environment.

    sudo apt update && \
sudo apt install -y python3 python3-pip python3-venv

    python3 -m venv venv && \
source venv/bin/activate

    Once the environment is ready, install boto3 with the following command:

    pip3 install --upgrade pip && \
pip3 install boto3==1.36.17

    Note: DL Streamer Pipeline Server expects the bucket to be already present in the database. The next step will help you create one.

  5. Create an S3 bucket using the following script.

    Update the HOST_IP and credentials with that of the running MinIO server. Use create_bucket.py as the file name.

    import boto3
url = "http://<HOST_IP>:30800"
user = "<value of MINIO_ACCESS_KEY used in helm/values.yaml>"
password = "<value of MINIO_SECRET_KEY used in helm/values.yaml>"
bucket_name = "ecgdemo"

client= boto3.client(
            "s3",
            endpoint_url=url,
            aws_access_key_id=user,
            aws_secret_access_key=password
)
client.create_bucket(Bucket=bucket_name)
buckets = client.list_buckets()
print("Buckets:", [b["Name"] for b in buckets.get("Buckets", [])])

    Run the above script to create the bucket.

    python3 create_bucket.py

  6. Start the pipeline with the following cURL command with <HOST_IP> set to system IP. Ensure to give the correct path to the model as seen below. This example starts an AI pipeline.

    Note: If you are running helm using an NGINX_HTTPS_PORT other than the default 30443, replace <HOST_IP> with <HOST_IP>:<NGINX_HTTPS_PORT>.

    curl -k https://<HOST_IP>:30443/api/pipelines/user_defined_pipelines/worker_safety_gear_detection_s3write -X POST -H 'Content-Type: application/json' -d '{
    "source": {
        "uri": "file:///home/pipeline-server/resources/videos/Safety_Full_Hat_and_Vest.avi",
        "type": "uri"
    },
    "destination": {
        "frame": {
            "type": "webrtc",
            "peer-id": "worker_safety_gear_detection_s3"
        }
    },
    "parameters": {
        "detection-properties": {
            "model": "/home/pipeline-server/resources/models/worker-safety-gear-detection/deployment/Detection/model/model.xml",
            "device": "CPU"
        }
    }
}'

  7. Go to MinIO console on https://<HOST_IP>:30443/minio/ and login with MINIO_ACCESS_KEY and MINIO_SECRET_KEY provided in helm/values.yaml file. After logging into console, you can go to ecgdemo bucket and check the frames stored.

    S3 minio image storage

  8. Uninstall the helm chart.

    helm uninstall app-deploy -n apps

MLOps using Model Download#

  1. Run all the steps mentioned in above section to setup the application.

  2. Install the helm chart

    helm install app-deploy helm -n apps --create-namespace

  3. Copy the resources such as video and model from local directory to the dlstreamer-pipeline-server pod to make them available for application while launching pipelines.

    # Below is an example for Worker safety gear detection. Please adjust the source path of models and videos appropriately for other sample applications.

POD_NAME=$(kubectl get pods -n apps -o jsonpath='{.items[*].metadata.name}' | tr ' ' '\n' | grep deployment-dlstreamer-pipeline-server | head -n 1)

kubectl cp resources/worker-safety-gear-detection/videos/Safety_Full_Hat_and_Vest.avi $POD_NAME:/home/pipeline-server/resources/videos/ -c dlstreamer-pipeline-server -n apps

kubectl cp resources/worker-safety-gear-detection/models/* $POD_NAME:/home/pipeline-server/resources/models/ -c dlstreamer-pipeline-server -n apps

  4. Modify the payload in helm/apps/worker-safety-gear-detection/payload.json to launch an instance for the MLOps pipeline

    [
    {
        "pipeline": "worker_safety_gear_detection_mlops",
        "payload":{
            "source": {
                "uri": "file:///home/pipeline-server/resources/videos/Safety_Full_Hat_and_Vest.avi",
                "type": "uri"
            },
            "destination": {
            "frame": {
                "type": "webrtc",
                "peer-id": "worker_safety"
            }
            },
            "parameters": {
                "detection-properties": {
                    "model": "/home/pipeline-server/resources/models/worker-safety-gear-detection/deployment/Detection/model/model.xml",
                    "device": "CPU"
                }
            }
        }
    }
]

  5. Start the pipeline with the above payload.

    ./sample_start.sh helm -p worker_safety_gear_detection_mlops

    Note the instance-id.

  6. Download and prepare the model.

    Note: For sake of simplicity, we assume that the new model has already been downloaded by Model Download microservice. The following curl command is only a simulation that just downloads the model. In production, however, they will be downloaded by the Model Download service.

    export MODEL_URL='https://github.com/open-edge-platform/edge-ai-resources/raw/06bb0d621cb14a1791672552a538beddddcc4066/models/INT8/worker-safety-gear-detection.zip'

curl -L "$MODEL_URL" -o "$(basename $MODEL_URL)"

unzip "$(basename $MODEL_URL)" -d new-model # downloaded model is now extracted to `new-model` directory.

  7. Run the following curl command to upload the local model.

    POD_NAME=$(kubectl get pods -n apps -o jsonpath='{.items[*].metadata.name}' | tr ' ' '\n' | grep deployment-dlstreamer-pipeline-server | head -n 1)

kubectl cp new-model $POD_NAME:/home/pipeline-server/resources/models/ -c dlstreamer-pipeline-server -n apps

  8. Stop the existing pipeline before restarting it with a new model. Use the instance-id generated from step 5.

    Note: If you are running helm using an NGINX_HTTPS_PORT other than the default 30443, replace <HOST_IP> with <HOST_IP>:<NGINX_HTTPS_PORT>.

    curl -k --location -X DELETE https://<HOST_IP>:30443/api/pipelines/{instance_id}

  9. Modify the payload in helm/apps/worker-safety-gear-detection/payload.json to launch an instance for the MLOps pipeline with this new model

    [
    {
        "pipeline": "worker_safety_gear_detection_mlops",
        "payload":{
            "source": {
                "uri": "file:///home/pipeline-server/resources/videos/Safety_Full_Hat_and_Vest.avi",
                "type": "uri"
            },
            "destination": {
            "frame": {
                "type": "webrtc",
                "peer-id": "worker_safety"
            }
            },
            "parameters": {
                "detection-properties": {
                    "model": "/home/pipeline-server/resources/models/new-model/deployment/Detection/model/model.xml",
                    "device": "CPU"
                }
            }
        }
    }
]

  10. Start the pipeline with the above payload.

    ./sample_start.sh helm -p worker_safety_gear_detection_mlops

  11. View the WebRTC streaming on https://<HOST_IP>:30443/mediamtx/<peer-str-id>/ by replacing <peer-str-id> with the value used in the original cURL command to start the pipeline.

Note: If you are running helm using an NGINX_HTTPS_PORT other than the default 30443, replace <HOST_IP> with <HOST_IP>:<NGINX_HTTPS_PORT>.

WebRTC streaming

Troubleshooting#