optimizing yield and size of potatos with stereovision and AI 

AVR wanted to develop a camera-based size measurement system to be installed on the bunker filling web of its Puma 4.0 harvester. The system was conceived to measure the length and width of potato samples in the product stream. 

1. Machine builders, not computer vision engineers

• Although the AVR R&D team had a clear idea of what the system needed to do, the company lacked the expertise and knowhow to develop it. That’s why AVR was looking for a reliable technology partner to work in close cooperation with the R&D team.

2. Finding the right partner

• “AVR not only wanted to develop a prototype, but also make the size measurement system production-ready,” says Pieterjan Vanlaeken, Platform Leader Technology at AVR. “This could only be achieved in collaboration with a partner with a strong affinity with manufacturing.”

AVR enlisted Vintecc to co-develop a size measurement system for the Puma 4.0 self-propelled potato harvester. The system uses a stereovision cameras mounted on the bunker filling web to determine the potatoes’ length and width during harvest. The benefit of stereovision is that this technology provides excellent depth estimation, which improves measurement accuracy. 

1. Size linked with location 

• The detection of potatoes is based on deep learning AI, which allows the system to recognize the potatoes in periodic sample images of the product stream. This in turn enables the system to select fully visible potatoes in the image for measuring their width and length.  
• The size measurement data can also be linked with geolocation data, which offers growers the possibility to evaluate width and length of the measured potatoes for  different harvested zones in their field. In other words, the AVR size measurement system provides a fully digital outlay of the width and length of the measured sample of the harvested potatoes. 

2. Computer vision in challenging conditions

• “Vintecc helped us to develop the system from the very start and really worked as an extension of our team,” says Pieterjan Vanlaeken. “They selected the technology and handled the AI model training. Especially the stereovision technology was new to us. Vintecc helped us to get the most out of this technology, even in challenging lighting conditions.”

3. Capture, an industrial IoT platform

• The first prototypes of the size measurement system were monitored by means of Vintecc's proprietary IoT platform Capture, which in this case collected valuable system data from the harvester in the field and made it available via the cloud for further evaluation by the R&D team.
• “Capture has been a great help during our first diagnosis and for debugging our first prototypes,” says Pieterjan Vanlaeken. “Without this IoT platform, we would never have been able to monitor the performance of our measurement system so efficiently.”

Training AI models typically requires lots of data. The challenge for the team was to obtain realistic potato images in large quantities and with a large variety relatively fast. 

1. Overcoming seasonal limitations

• Ideally, the image data needed to come from various types of potatoes in different fields, in different lighting conditions, and on different types of conveyor belts. But as potato harvesting is a seasonal activity, limited to a few months of the year, it was highly unpractical to obtain this data in a realistic time frame.

2. 3D rendering for faster AI training

• Vintecc therefore decided to train the AI model with synthetically generated data. That is data created through 3D rendering and by simulating various realistic lighting and environmental conditions. The synthetic data managed to render very realistic textures and details of potatoes in various perspectives and conditions.

3. Reduce time-to-market

• “The use of synthetic data helped us to significantly reduce our time-to-market,” says Pieterjan Vanlaeken. “Working with real-life data alone would never have worked for us, and Vintecc was instrumental in directing us towards the much faster path of synthetic data. Eventually, to further optimize our results, we did mix synthetic with real data.”

The new size measurement system for the Puma 4.0 harvester was presented to the public at Interpom 2024, Europe’s leading indoor trade event for the potato chain. At that time, the first harvesters equipped with the final prototypes of the new AVR technology had already been subjected to extensive field testing.