stream-object-detection

Listen to webcam feed and perform object detection

---

Getting started:

1. Overview
2. Installation
3. Methods
4. K8s

---

Overview

An application that uses pre-trained PyTorch model to perform object detection in real-time where the video feed is acquired from webcam, packaged using docker and prepared for deployment on Kubernetes

The whole process is completes in 2 steps:

1. Frames are received from the webcam.
2. The frames are processed using the pre-trained model (yolov5s)
3. If person is detected with confidence greater than 0.5, we consider it as anomaly and send a notification.

The main files and folders in the directory are:

1. webcam
   • 📑 app.py - contains the APIs.
   • 📑 Dockerfile - used for building the docker image.
   • 📑 requirements.txt - used for installing the dependencies.
2. live_stream
   • 📑 client.py - Client side code to stream the video feed from webcam to the server.
   • 📑 server.py - Server side code to process the frames and send the anomaly notifications.
   • 📑 Dockerfile - used for building the docker image.
   • 📑 requirements.txt - used for installing the dependencies.

---

Installation

• The Docker image can be built using the following command:

  docker build -t sanidhya7921/stream-object-detection .
  

• To test your Docker image locally.:

  docker run sanidhya7921/stream-object-detection     
  

ℹ : to push your Docker image to Docker Hub, use the following command-
docker push sanidhya7921/stream-object-detection

---

Methods

There are 3 main classes:

Class Name	API	Description
WebCam	WebCam.get_frame	Get the frame from the webcam
Inference	Inference.predict	Predict the objects in the frame using yolov5s model
AnomalyDetector	AnomalyDetector.detect	Detect if the results are anomaly or not

---

K8s Deployment

Deployment resource:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: stream-object-detection
spec:
  selector:
    matchLabels:
      app: stream-object-detection
  replicas: 5
  template:
    metadata:
      labels:
        app: stream-object-detection
    spec:
      containers:
      - name: stream-object-detection
        image: stream-object-detection
        imagePullPolicy: Always
        ports:
        - containerPort: 5000

---

Service resource:

apiVersion: v1
kind: Service
metadata:
  name: stream-object-detection
spec:
  selector:
    app: stream-object-detection
  ports:
  - protocol: "TCP"
    port: 5555
    targetPort: 80
  type: LoadBalancer

HPA Resource:

apiVersion: autoscaling/v1
kind: HorizontalPodAutoscaler
metadata:
  creationTimestamp: null
  name: stream-object-detection
spec:
  maxReplicas: 2
  minReplicas: 1
  scaleTargetRef:
      apiVersion: apps/v1
      kind: Deployment
      name: stream-object-detection
  targetCPUUtilizationPercentage: 70

ConfigMap resource:

If we want to use .env.prod the following command can be run in the terminal of lens.

Kubectl create configmap <map_name> --from-file=".env.prod"

---