After years of continuous development in the past decades, the field of forest production has achieved a notable degree of mechanization. Nonetheless, the unceasing demand for increased productivity in forestry operations continues to drive the rapid expansion of forestry automation. Critical aspects of forest production, including activities such as harvesting, forwarding, seedling planting, and soil preparation, hold significant potential for improved automation. These advancements promise numerous advantages, such as heightened production efficiency, reduced environmental impact, reduced operational costs, and the creation of a more favourable work environment. Beyond fostering industrial progress, such innovations could also make a positive contribution to the sustainable development of forestry.

Forestry operations take place in intricate natural forest environments, rendering the automation of these processes complex and challenging. To realize full automation of forest machines, it is imperative to equip them with environmental awareness and cognitive capabilities. This can be achieved by integrating advanced imaging sensors with sophisticated algorithms, thereby enabling forest machinery to possess a vision system. By utilizing both novel and existing solutions for object detection, positioning, categorization, and status analysis within the machine's surroundings, combined with intelligent decision-making and control mechanisms, forest machinery can attain a higher degree of operational automation.

This thesis primarily focuses on the development, deployment, and validation of vision systems on an unmanned forestry machine platform, with a special emphasis on the prototype development of essential features that are still manually executed in contemporary forestry operations. These features include forest terrain obstacle detection, roundwood pose estimation, automated selection of seedling planting locations, and autonomous obstacle avoidance for mounders.

In this thesis, vision systems rooted in color and stereo camera sensing are designed and deployed on an unmanned forest machinery platform. These systems encompass the following key functionalities:

•             A precise positioning system for detecting forest terrain obstacles, such as stones and stumps, using stereo camera data in conjunction with deep learning techniques.

•             Localization and pose estimation capabilities for ground logs, leveraging stereo cameras and deep learning.

•             An analysis of the planning area for obstacle detection, along with the extraction of feasible planting zones for establishing seedling planting locations.

•             The development of an automated obstacle avoidance system for forestry mounders, powered by visual solutions.

Through rigorous testing in real-world scenarios spanning logging, loading, planting, and site preparation, this paper demonstrates the feasibility and practicality of enhancing the level of automation in forest machinery operations through the integration of vision systems. It culminates in the creation of tangible and efficient functional prototypes, leading towards a new era of automation in the field of forestry.