Toward Smart Forestry: A Review on Vision-based Perception for Terrain Classification and Obstacle Detection

Ground autonomous systems are increasingly being explored for forestry applications, offering potential to improve safety, efficiency, and scalability in operations such as harvesting, monitoring, and transportation. A fundamental component enabling this autonomy is perception, which allows machines to understand and react to complex environments. This paper presents a comprehensive review of vision-based perception techniques for forestry, with a particular focus on terrain classification and obstacle detection. For terrain classification, we examine four areas: learning-based approaches for terrain understanding; fusion of visual and proprioceptive sensory information for enhanced classification accuracy; terrain property estimation techniques for assessing surface characteristics such as slope and traversability; and vegetation classification methods for distinguishing plant structures that impact navigation decisions. In obstacle detection, we systematically analyze five critical categories: terrain-related obstacles including rocks, roots, and uneven surfaces; vegetation-based obstacles such as fallen trees and dense undergrowth; wild animals that pose dynamic threats to autonomous operations; environmental obstacles including lighting variations, fire/smoke, and weather conditions; and debris and ground-level obstacles such as scattered branches, stumps, and human-made objects. Covering literature from 2019 to 2025, this review highlights the current state-of-the-art, identifies key limitations, and outlines future research directions for deploying perception systems in real-world forestry scenarios.