Automatic Growth Tracking System for Medical Plants

About the Client

A pharmaceutical company that is operating automated plantations of medicinal plants.

Business Challenge

On his plantations, our client experiments with various agricutechnical schemes (a different scheme for each bush).

Currently, the customer uses robots with cameras for regular plants observation.

But the business needed more automation and modernization. To reduce the human factor and achieve complete market security, it was suggested to make the client’s business more automatic and modern.

Also, it’s essential to control and monitor tops growth process and anomalies on the top level for the schemes of further analysis and enhancement.

The client needs to monitor the tops growth process and anomalies to use the information for further analysis and enhancement.

Solution Overview

Quantum used robots with cameras for periodical plant observation.

There are images from two different cameras that came every hour.

  1. To better show differences and get the opportunity to track changes operatively, we use a one-week interval between images.
  2. It takes approximately 20 seconds per image to analyze and detect many tops with an embedded device on one image.
  3. As a result, we got 90% of accuracy.

Also, we created a computer vision solution that takes images coming from different zones with some time intervals.

It detects growing tops and identifies them, providing information on the growth process as a result of each unique growing top.

 

desktop

Project Description

During this project, we used the CRISP-DM methodology.

Business understanding

Our team explored the process of medicinal plant growth.

Data understanding

We had images from two different cameras that came every hour and used them as an input. The team had to decide what interval between the photos would show us differences and let us track changes at the same time. We went with a one-week interval between images.

Data preparation

At this stage, we prepared data for future manipulations. This included manually labeling images as well as converting them to a lower resolution. By doing so, we made the processing faster while keeping accuracy.

Modeling

The Quantum team decided to choose YOLO as the fastest and the most popular way to get high-level accuracy and fast processing. We spent approximately 12 hours on transfer learning for 30,000 iterations.

Tops tracking was a challenge. Plants don’t usually grow strictly up and can change their growth direction. We had to identify the same tops over time, so, we built a map of tops that helped us identify the necessary one by dividing a shot into sectors and finding the top that was closest to the previous shot.

Evaluation

In the end, we got 90% of accuracy. It takes approximately 20 seconds per image to analyze and detect many tops with an embedded device on one image.

Deployment

Quantum deployed the solution on Jetson TX2.

Let's discuss your idea!

image-2

Technological Details

The data was labeled using the LabelImg tool. After getting bounding boxes from the tool, the YOLO model was trained to detect grow tops.

We implemented a custom Python algorithm that used a special distances vector to describe every grow top on the image.

For bud sizes, we tuned an OpenCV bbDetector. In the end, the system tracked the growth process of every top of the plants.

Python
Python
OpenCV
OpenCV
Numpy
Numpy
nvidia-docker
nvidia-docker

Connect with our experts