Determination of the size of rock granules in a bucket and on a conveyor

Computer vision
Сhallenge
01/
Very small pellets indicate that extra explosives are spent on drilling and blasting - it's expensive.
The size of the rock pellets is an indicator of the efficiency of the entire mining process.
Too large granules reduce the performance of the bucket, because voids form in it. And they also require more processing time and can even break the crushing machine.
Counting granules of different sizes using computer vision algorithms.
Solution
It is necessary to determine the number and size of granules in the bucket and on the belt and find granules of too large size.
Task
Why is it useful
02/
automation of the verification process
Employees do not have to evaluate the composition of the conveyor belt by eye – and this is additional time and qualifications:
acceleration and optimization of bucket/belt operation
How it works
03/
The data about this is transmitted to the customer
The system analyzes the video and highlights granules of too large size
Cameras are installed on the conveyor and bucket (if they are not already there)
The bucket is dark
Additional lighting and additional training of the model were needed to work in difficult visual conditions.
Solution
If the video quality was sufficient indoors, then it was more difficult to work "in the fields". The bucket is quite dark during the day, and often the work goes on at night.
Problem
Large mass of similar objects
Therefore, markup played an important role in the project — and we had to work with it.
Solution
There could be from 600 to 1000 objects in one picture, and they all look pretty similar. Besides, the conveyor belt is moving fast.
Problem
The problems we are facing
04/

Project team

05/
Maxim Lukin
Middle software engineer / DS & ML engineer
Tech lead
Nikita Buzanov
Senior software engineer
Sergey Solovyov

Tools

06/
zDialog фреймворк

Detectron2

Library for object recognition.
OneDash сервис
More

Markup

Internal markup tool

Rendering by granules

07/

Dumpcars/belaz

08/

What we have learned

09/
Kafka
calibration of cameras
HLS-stream
working with 3D cameras and depth maps
Contact us