As pressures increase on the food supply chain, the importance of predictable yields increases. Generally, the better informed growers are, the better the price they can potentially obtain. However, manual counts struggle to achieve sufficient accuracy. Although counting fruits or salad is a simple task for a human, it has proven hard to automate.

We help our customers improve end-to-end efficiency through the integration of technology.

Enter artificial intelligence – these new techniques enable applications which weren’t possible before, such as identifying individuals or autonomous driving. These systems can give amazing results – but they rely on either high cost, fragile hardware locally or having a fast Internet connection to the machine. If AI on the farm is to be worth the investment, it needs to be much cheaper and mobile.

The challenge is to squeeze the power of a high-end desktop machine into a small electronic module which could be put on the back of a tractor. Well, there is an industry trend which benefits us – moving compute away from data centres to the ‘edge’, putting AI into places like traffic cameras, Internet of Things (IoT) devices and domestic appliances. As result, silicon vendors are investing heavily into making processors which are AI-capable but cheap and rugged. Also, the task of recognising fruit to count them is simpler than many AI tasks, which reduces the effort needed to train and run the network.

This video shows an example application: counting apples from a live camera stream. It’s a classic example of something which is hard to program using traditional techniques: apples have variable sizes and colours, which would take continual adjustment of the settings to maintain accuracy. However, the AI system can be trained using example pictures, which allows it to cover the natural range of appearances without becoming ‘snap happy’: catching everything which is round or green.

The ruggedness and low power requirements of this design (based on a standard single board computer) mean that it’s suitable for battery powered machines, such as autonomous robots. The fact that it doesn’t need an internet connection also means that it runs in real time – which enables individual treatment of plants or robotic harvesting.

The technology developed to make industries and cities smarter and more efficient can benefit food production as well – helping farmers get better value from their land in a changing world.

 

Author
Simon Jordan
Senior Sensor Physicist

Working in our sensing systems group, Simon specialises in navigation and communication. Before joining Cambridge Consultants, he spent ten years at Teledyne TSS, working on projects including electromagnetic pipe tracking/survey systems, ship steering systems, marine motion sensors, and the development of high grade inertial navigation systems.