Counting on computer vision
By providing more precise yield forecasts, the computer vision system could help growers allocate human and financial resources more efficiently while planning harvesting and distribution logistics more effectively.
By Ron van der Ploeg
A computer vision-based plant counting system developed by researchers and students at the University of Antioquia (UdeA) is being fine-tuned for integration into the production operations of Flores El Capiro, a large-scale chrysanthemum grower located in Colombia’s northwestern department of Antioquia.
Chrysanthemum harvesting in Colombia. More reliable plant counts could help growers plan harvesting and distribution. Photo: Floritec.
Counting the crop
Sicop: more accurate harvest projections Known as Sicop, the system was developed by the university’s Power Electronics, Automation and Robotics Group (Gepar) to improve the counting of chrysanthemum young plants in growing beds.
The project emerged from collaboration between UdeA and Flores El Capiro, combining the company’s production needs with expertise developed within the university. At its core, Sicop uses computer vision, a technology that enables computers to identify and interpret visual information in a way that resembles aspects of human vision. It allows machines to analyse images, recognise objects, detect patterns and extract information from visual inputs using a range of image-processing algorithms, including techniques for identifying shapes and detecting edges.
The system establishes a standardised approach for capturing video of chrysanthemum beds containing young plants. The accompanying software includes two main modules: one creates a visual overview of the crop, while the other processes and analyses the captured information.
According to the system’s user manual, Sicop is a practical application of research into computer vision techniques. Its intended purpose is to provide more reliable information about the number of young plants in each bed, ultimately supporting more accurate harvest projections. Addressing a difficult counting challenge The need for the technology stems from the scale and complexity of Flores El Capiro’s production sites.
Professor David Stephen Fernández McCann, coordinator of Gepar, explained that the company’s crop technicians oversee and inspect more than 1,000 beds each week, containing approximately 4.5 million plants. Because planting is carried out manually and the distribution of plants can vary considerably, obtaining an accurate count is a challenging task.
Existing commercial applications had already demonstrated the potential of automated plant counting, but Flores El Capiro found that they were not sufficiently accurate for its particular growing conditions.
José Manuel Henao Londoño, commercial manager at Flores El Capiro, stressed that conventional applications can mistake leaves or other visual elements for individual plants.
“In our case, the plants come in different shades of green; if a leaf falls off, these apps count it as another plant, leading to errors in the exact count,” he said.
This prompted the company and university researchers to pursue a more specialised technological solution that could account for the particular characteristics of the crop.
The drone option was quickly discarded because of physical obstacles present in the growing areas, including wires, sticks and lighting fixtures Photo credit: Universidad de Antioquia.
In our case, the plants come in different shades of green; if a leaf falls off, these apps count it as another plant, leading to errors in the exact count
José Manuel Henao Londoño Flores El Capiro
Testing in practice
Developing the right imaging method
The team tested several approaches before settling on the configuration used for Sicop.
Initial experiments involved a handheld GoPro camera, with an operator walking along the entire path and taking photographs of the beds. Researchers also considered using a drone to capture images from above.
However, the drone option was quickly discarded because of physical obstacles present in the growing areas, including wires, sticks and lighting fixtures.
After multiple trials, the researchers settled on the handheld camera configuration. The development process demonstrated the importance of adapting computer vision technology to real agricultural conditions rather than relying solely on existing commercial solutions.
University-industry collaboration
Karla Cristina Bonilla Restrepo, from UdeA’s Innovation Division, sees the project as a powerful example of cross-sector collaboration, demonstrating how partnerships between academia and industry can translate research into practical solutions.
She said: “Companies operate in the real world, and the University may not be fully aware of their needs. Through coordinated efforts, we can better identify those needs and provide timely solutions.”
A smartphone mounted on a handheld stabiliser above young chrysanthemum plants. Photo: Universidad de Antioquia.
Towards routine use
Refining the system
The project was completed in February 2021. According to Henao Londoño, Flores El Capiro is currently in the stage of refinement, which involves polishing and perfecting the tool so that it resonates with users and stands up to routine application.
Some factors can affect the plant-counting result. Take, for example, the overlapping foliage as the seedlings develop. These may cause difficulties in distinguishing one plant from the next. There is also the presence of weeds, as discriminating these from the actual seedlings may be impossible. Another issue is variation in light levels inside a greenhouse, which may distort captured images and lead to incorrect counting.
In addition, the system does not detect pest or disease symptoms or presence; this could be considered for future improvements and would certainly add to the usefulness of the system.
Potential beyond a single company
The researchers see broader potential for Sicop because the challenge it addresses is not unique to Flores El Capiro. Other flower and agricultural producers also need reliable information about plant populations within their growing beds, while manual counting can require significant time and effort.
However, the system is purpose-built for chrysanthemum starter plants. So, using it with a different crop may lead to reduced accuracy. Recalibration or adjustments for other crops – including field-grown cut flowers – could be considered in future to expand its application and usefulness.
Bonilla said that projects of this type can be difficult for some organisations to finance, making institutional and public support particularly important. In this case, the project benefited from public funding through the university, allowing the development of a prototype based on a real industry requirement.
Beyond its immediate application, the prototype could help companies strengthen their production capabilities and modernise processes that continue to rely heavily on manual counting.
