Computed vision with WaziGate
In this course, we will learn how to build a WaziApp capable or doing object detection from a webcam. With the help of this application you can utilize the WaziGate and RaspiCam to track objects and evaluate their occurrences.
The models are trained with the COCO (Common Objects in Context)  dataset. It contains 80 classes, you can have a look on them here.
From now you can choose between two models for object detection.
- Mobile Net v1 (224x224) 
- Yolo Version 5s (320x320) 
The App is in an early stage of development, it is still work in progress, some features are missing.
You can just download the build docker image from the Waziup Dockerhub directly on your RPI.
docker pull waziup/wazigate-webcam-inference
The WaziApp is available at github
If you want to build it from source you have to follow these steps.
- Download the repository to your local machine:
git clone https://github.com/Waziup/wazigate-webcam-inference.git
- Install docker and issue the following commands
docker buildx create --name rpibuilder --platform linux/arm/v7 docker buildx use rpibuilder docker buildx inspect –bootstrap docker run --rm --privileged multiarch/qemu-user-static --reset -p yes
- Navigate to the repository:
- Issue the following command to build the docker image from the Dockerfile:
docker buildx build –platform linux/arm/v7 -t waziup/wazigate-webcam-inference:latest –load .
- To copy the image via SSH to the raspberry pi with the following command:
docker save <id_of_build_image> | gzip | pv | ssh pi@<ip_of_pi_in_local_network> docker load
- It can occur that the name of the repository is lost, so tag the image appropriate to the docker-compose.yml
docker tag <id_of_build_image> waziup/wazigate-webcam-inference:latest
- Afterwards, start the application with via the UI of the WaziGate or run the following command, to see the logs in console:
Your WaziApp is now up an running!
For now, like mentioned above, there are two networks someone can choose from. In the following there is some information about the used networks.
MobileNet is a simple convolutional neural network it has 30 layers and 4.2 million parameters. In the following there is an overview of the architecture.
Further information about the network is given in the paper of the inventors.
Yolov5 is a state of the art realtime object detection network . It has got 213 layers, 7.2 million parameters. In the following there is an overview of the architecture.
Further information about the network is given here. There is no Yolov5 paper, to this date, but Yolov4 has a paper and does not differ much.