Kauricone – Building Waste Machine Learning Solution

 

 

Solution

 

Kauricone IoT Server uses Machine Learning to collect data from Cameras, and monitors the types and amounts of waste going into Skips on Building Sites. The objective is to identify the building waste, and then begin to identify ways of reducing it

 

The Kauricone Machine Learning Process

Image Recognition

Kauricone are focussed on Image Recognition applications for Machine Learning. To do this we follow this process:

1. Collect more than 2000 images of the object to be detected

2. Annotate each image (ie put a border around each object to clearly identify it from any other object in the same image frame)

3. Identify as many traits of the object as possible

4. Begin training the machine with the objects, and traits that are available

5. Take images of the objects in the production environment

6. Process these images against the trained model at regular intervals

7. Look for exceptions and trends which require action

Predictions

This is the output from the server, after an image has been processed

Requirements

1. Kauricone IoT Server (4GB, 128 eMMC Storage, ARM 6 Core Processor, Connection Interfaces)

2. Camera (Wifi, 4G, Network, USB)

3. Computer

4. Ubuntu 18.04, Tensorflow, Fast CNN, Python (Preinstalled)

Process

 

Camera takes pictures

of Skip, Motion Activated

Solar, 4G

 

 

 

Camera connects to Kauricone

IoT Server

by 4G via Cellular Network

 

 

 

 

Kauricoe IOT Server uses

machine learning to process the Data from the Camera to identify

the Building Waste by type

eg Plaster board, Framing Timber, coffee cups

 

 

 

Contents of Waste in Skip

is Monitored, via any PC or

Smart Phone

 

 

The Camera is mounted above a skip bin on a building site. The camera is Solar Powered, motion activated, and sends images via the 4G cellular network.

 

 

 

 

 

 

 

 

Skip Bin filling up

 

 

 

 

 

 

 

 

 

 

These images represent the steady filling up of the skip bin, and to date the Skip Bin has been emptied 3 times. An average day results in images being sent from 1 to 20 times depending on how many times the camera is activated. We have found we get irrelevant motion activations from vehicles passing by, or sheets of plastic being blown in high winds

 

 

Machine Learning Output