WP8 is devoted to sensory and perception solutions for forestry automation applications. The proposed tasks are object detection/classification (bushes, rocks and trees) and estimation of ground bearing capacity (for propulsion of both manual and autonomous forest machines). The developed solutions will also be applied to robots in vineyard environments (WP7) and greenhouse environments (WP5).
In Europe today the Cut-To-Length (CTL) system is used in forestry. The system utilizes two machines: a harvester and a forwarder. The harvester fells the trees, debranches them and finally crosscuts them in the desired length. This is a highly automated process with just a little intervention from the operator to get as high productivity as possible; a modern harvester can handle up to 100 trees per hour. The forwarder then picks up the logs from the ground and transports them to a roadside landing for further transportation to the saw-mill.
A harvester (left) fells the trees, debranches them and finally crosscuts them in the desired length. A forwarder (right) transport the logs to a nearby road for further transport to the saw-mill
The forest industry is under keen competition, and therefore must constantly develop its products to make them more profitable. Since the modern harvester is able to make use of almost 100% of the trees' value, the harvesting speed (m3/h) must be increased to make it more effective. A problem with this is that the operator works under stressful conditions today, and if the harvesting speed rises even further the operator risks becoming a bottleneck. To solve this, the operator would have to be replaced by a fully autonomous machine, or at least get help from semi-autonomous systems. The research tasks in WP8 deal with two important requirements for autonomous navigation; techniques to detect and classify objects, and techniques to estimate ground baring capacity.
Ground bearing capacity
The most common reasons for ground damage and wheel rutting in forestry are the bearing capacity of the ground and the ground pressure of the forest machines. The expected more rainy periods and less frozen ground in northern Europe due to climate change will increase soil moisture content and reduce the bearing capacity which in turn would increase wheel slippage. Good planning before harvesting should steer the operation towards better areas, but unexpected heavy rains can alter ground conditions very fast. Heavy axle loads and wheel slippage will create deep rutting and soil compaction and increase the fuel consumption. This can in turn lead to less growth of nearby trees and reduced biological diversity. Thus, there is generally a conflict between minimized soil disturbance and maximized operational efficiency. In practice, the operational efficiency often takes precedence over environmental concerns. One of the few financial incitements for reducing ground damage is the risk of being fined by the EU (Water Framework Directive).
Two things we try to avoid by measuring ground bearing capacity: the machine getting stuck (left) and wheel rutting (right).
Since soil type and soil moisture contents is one of the key factors determining ground bearing capacity, a sensor able to measure moisture contents would be highly useful on both manual and autonomous forestry machines. Given knowledge about current soil type, a moisture sensor would make it possible for an operator to avoid areas with low bearing capacity. Future autonomous systems could use the same type of sensor for short-term path planning, thus reducing ground damages and the risk to get stuck due to too low bearing capacity. In this scenario, GIS maps in combination with GPS could provide information on current soil type.
Object detection and classification
For autonomous forestry vehicles, the ability to detect and classify objects in the vicinity of the vehicle is crucial. The vehicle must be able to identify obstacles such as trees, stumps, stones, and holes in the ground. This serves as important information for safe navigation through the forest. The system has to distinguish between an object that the vehicle has to avoid and one that it can safely run over, e.g. a rock versus a bush. This information can be used as input to a map building system and also as input to an emergency stop function. The vehicle must also be able to detect human beings appearing dangerously close to the vehicle. Such detection would, depending on the estimated range, result in temporary suspension of operation, or an emergency stop action.
Human detection using thermal camera
For safe operation of autonomous forest machines there is a need to detect and avoid humans nearby the vehicle. In the Crops project we have developed an algorithm that can detect humans in a forest environment by analyzing temperature differences in images from a thermal camera. The algorithm works by first locating objects in the image with higher temperature than the surrounding. Then a validation algorithm is applied to determine if an objects is a human. When applied to 57 images with 94 objects (humans, cars, trees, buildings), the precision was 98.6% and the false discovery rate was 1.4%. The video below shows the system tracking two people walking in a dense forest.