Automatic Data Labeling Can Make Forestry Safer

Researchers at Luleå University of Technology have developed methods that allow computers to automatically generate training data for perception systems in autonomous work vehicles. The approach can reduce manual labor, improve precision, and ultimately contribute to a safer and more sustainable forestry industry.

Self-driving vehicles are becoming increasingly common on public roads, and research is now underway to increase the level of automation for off-road vehicles in sectors such as forestry and construction. These vehicles can take over hazardous tasks, but for that to work, they must be able to perceive their surroundings with high accuracy. Training data plays a crucial role here, and researchers at Luleå University of Technology have found a way to let computers generate it automatically.

"This allows us to free up time, reduce risks for humans, and at the same time pave the way for a more environmentally friendly and resource-efficient industry," says Magnus Karlberg, Professor of Machine Design at Luleå University of Technology.

Magnus Karlberg, professor of machine design at Luleå University of Technology.

Virtual Training Improves Precision

The project, funded by the Jubilee Fund at Luleå University of Technology, shows that it is possible to train models using automatically labeled data (also known as image annotation) generated in a virtual environment. When a smaller amount of real data is combined with synthetic images, accuracy improves even with a limited training set.

The largest errors often occur for logs located far away or in shaded areas, pointing to clear directions for continued research. A case study also shows that the stem volume of standing trees can be estimated automatically using a low-cost stereo camera combined with publicly available YOLO models.

"Our tests show that a small amount of real data, when combined with synthetic images, can produce better accuracy. The next step is to expand the datasets with greater environmental variation and build more virtual models of forest objects," says Magnus Karlberg.

From Game Engine to Research Data

To create the synthetic images, the researchers recreated a physical test environment in the Unity game engine. A digital twin of the AORO platform — an off-road research platform within the Arctic Off-Road Robotics Lab used to develop autonomous vehicles for future forestry — was connected to the environment and equipped with an emulated stereo camera. This made it possible to generate automatically labeled images that closely mimic real driving situations.

For smaller objects, such as forest seedlings, the team used photogrammetry, where a robot moved the camera between different positions. The raw data was converted into 3D models, enabling large amounts of training data to be produced with minimal manual effort.

"The virtual workflow opens up new possibilities. We can test, adjust, and improve models in the computer before anything is built in reality, which saves both time and resources," says Magnus Karlberg.

The research results were presented at the ISTVS conference in Yokohama in October 2024 and have been selected for extended publication in the Journal of Terramechanics.

• Machine design