Object Detection at the Edge

Prerequisites

Install the following tools on your Edge device: Docker. Use the following links to help you:

Docker Community Edition installation

My Edge device is currently running Fedora 27, with the following versions of the abovementionned packages:

Docker version 18.02.0-ce, build fc4de44

We use IP cameras from SV3C, more specifically the POE models (that can do 1080p): http://www.sv3c.com/POE-IP-Camera.html

Getting ready

Clone this repository locally: git clone https://github.com/inc0/video_detection
Make sure Docker is running: sudo systemctl start docker
Put yourself in the cloned folder: cd video_detection

Running the Object Detection algorithm in containers

Setting up Kubernetes

On Fedora 27 Sources of inspiration (aka online guides):

sudo dnf install kubernetes-client kubernetes-node kubernetes-master kubernetes-kubeadm

That installs Kubernetes 1.9.1 on your Fedora 27 system.

Modify the kubelet Cgroup driver to use the same as Docker:

$ sudo docker info | grep -i cgroup
$ sed -i "s/cgroup-driver=systemd/cgroup-driver=cgroupfs/g" /etc/systemd/system/kubelet.service.d/kubeadm.conf

Note: we will reload the daemon and restart the kubelet service later on (see below)

Other preparation tasks:

$ sudo swapoff -a
$ sudo systemctl disable firewalld
$ sudo systemctl stop firewalld
$ sudo systemctl enable kubelet.service

Note: using swapoff does not turn off the swap permanently, it will be active again after a reboot. If you wish to make this change permanent, do this:

sudo sed -i "s/\/dev\/mapper\/fedora-swap/#\/dev\/mapper\/fedora-swap/g" /etc/fstab

Make sure docker is using runc and fix the CNI bin path used:

$ sudo docker info | grep Run
$ sudo sed -i "s/--cni-bin-dir=\/usr\/libexec\/cni/--cni-bin-dir=\/usr\/libexec\/cni --cni-bin-dir=\/opt\/cni\/bin/g" /etc/systemd/system/kubelet.service.d/kubeadm.conf

Restart the kubelet service

$ sudo systemctl daemon-reload
$ sudo systemctl restart kubelet

Note: running sudo kubeadm init on my system takes the network interface down... not sure why, under investigation. [Solved]: Docker was using cc-runtime by default... it worked after changing this back to runc.

Create the Cluster

$ sudo kubeadm init --pod-network-cidr=192.168.150.0/24
$ cd k8s && wget https://docs.projectcalico.org/v3.0/getting-started/kubernetes/installation/hosted/kubeadm/1.7/calico.yaml && cd ..
$ sed -i "s/192.168.0.0\/16/192.168.150.0\/24/g" k8s/calico.yaml
$ kubectl apply -f k8s/calico.yaml

Note: I really don't know what valid or not in terms of --pod-network-cidr, I'm mostly trying to avoid any clash with other similar subnets I have locally.

Important: write down the kubeadm join command that will be printed out at the end of the initialisation phase. It will be used to get nodes to join the cluster.

Allow non-root users to use kubectl (optional)

$ mkdir -p $HOME/.kube
$ sudo cp -i /etc/kubernetes/admin.conf $HOME/.kube/config
$ sudo chown $(id -u):$(id -g) $HOME/.kube/config

Verify that all containers are running:

$ kubectl get pods --all-namespaces

Allow pods to be run on your master (optional)

kubectl taint nodes --all node-role.kubernetes.io/master-

On Clear Linux (using Clear Containers)

$ sudo swupd bundle-add cloud-native-basic

$ sudo swapoff -a

Note: you can turn the swap off permanently by ???.

You may get this error when trying to join the cluster network:

	[ERROR CRI]: unable to check if the container runtime at "/var/run/dockershim.sock" is running: exit status 1

If that's the case, add --ignore-preflight-errors=cri to the kubeadm join <blah blah> command-line.

To be completed - still having some networking issues

Running the Object and Zombies detection algorithms

This section assumes that you have a working configuration based on Kubernetes. The steps above are not complete in that regard.

In order to run the demo, here are the steps to follow (at a high-level):

Build the containers you will need
Start the private registry
Upload the containers (in case they are not yet available in the private registry
Start the various services
Visualize the output
Shutting down the demo (and restarting it)

Because this document would become too long with this, I have moved this section in a dedicated file named running-the-demo.md

Using Docker Compose

Install

Follow these on-line instructions: Docker Compose installation. I have tested this on Fedora 27 using this version of docker-compose:

docker-compose version 1.19.0, build 9e633ef

Note: you do not need to need to build any Docker container. It will be built automatically for you if not yet available. See the build instruction in the docker-compose.yml file.

Running the Object Detection algorithm

On Fedora 27 Note: unless you have installed Docker and performed these post-installation steps, you will need to use sudo to use docker-compose. For some reason, it does not find the docker-compose executable and I have to use the full path to it, i.e. sudo /usr/local/bin/docker-compose up (to be investigated later, or maybe a reboot will solve it).
Set-up the IP camera following the online User Guides
Take note of the IP address assigned to the camera

We use an environment variable to pass the IP address of the camera to the container that will capture the video stream. The name of the variable is IP_CAMERA. 3. Start the Object Detection service

sudo IP_CAMERA=192.168.0.130 /usr/local/bin/docker-compose up

Open a broswer at http://0.0.0.0:5000 to see the results
On Clear Linux It is possible to run the same set-up on Clear Linux. Unfortunately, it is not possible to use Clear Containers for this yet as name resolution does not work in either Docker Compose or Docker Swarm. See the "Name resolution does not work when using docker-compose" issue for more details.

Same as for Fedora 27 (or any other OS really), install Docker and docker-compose. Use the following guides to help you:

Installing Clear Containers 3.0 on Clear Linux
Install docker-compose:

$ sudo curl -L https://github.com/docker/compose/releases/download/1.19.0/docker-compose-`uname -s`-`uname -m` -o /usr/bin/docker-compose
$ sudo chmod +x /usr/bin/docker-compose

Because of the current issue with Clear Containers 3, we have to change the default behaviour of Docker when running in Clear Linux (bare metal). The background is that, by default, it will pick the Clear Containers 3 runtime (cc-runtime) when running bare-metal so we need to force the system to not do that.

One way to achieve this (not sure whether it's the best way) is to modify the docker.service file (/lib/systemd/system/docker.service) and change the ExecStart line to this:

ExecStart=/usr/bin/dockerd --storage-driver=overlay2 --default-runtime=runc

Now reload and restart the daemon:

$ sudo systemctl daemon-reload
$ sudo systemctl restart docker

Verify that the runc runtime is being used:

sudo docker info | grep Run

You are now ready to continue from Running the Object Detection algorithm

gvancuts / video_detection