TY  - THES
T1  - Perception for context awareness of agricultural robots
A1  - Reiser,David
Y1  - 2019/03/06
N2  - Context awareness is one key point for the realisation of robust autonomous systems in unstructured environments like agriculture. Robots need a precise description of their environment so that tasks could be planned and executed correctly. When using a robot system in a controlled, not changing environment, the programmer maybe could model all possible circumstances to get the system reliable. However, the situation gets more complex when the environment and the objects are changing their shape, position or behaviour. Perception for context awareness in agriculture means to detect and classify objects of interest in the environment correctly and react to them. 
The aim of this cumulative dissertation was to apply different strategies to increase context awareness with perception in mobile robots in agriculture. The objectives of this thesis were to address five aspects of environment perception: (I) test static local sensor communication with a mobile vehicle, (II) detect unstructured objects in a controlled environment, (III) describe the influence of growth stage to algorithm outcomes, (IV) use the gained sensor information to detect single plants and (V) improve the robustness of algorithms under noisy conditions.
First, the communication between a static Wireless Sensor Network and a mobile robot was investigated. The wireless sensor nodes were able to send local data from sensors attached to the systems. The sensors were placed in a vineyard and the robot followed automatically the row structure to receive the data. It was possible to localize the single nodes just with the exact robot position and the attenuation model of the received signal strength with triangulation. The precision was 0.6 m and more precise than a provided differential global navigation satellite system signal.
The second research area focused on the detection of unstructured objects in point clouds. Therefore, a low-cost sonar sensor was attached to a 3D-frame with millimetre level accuracy to exactly localize the sensor position. With the sensor position and the sensor reading, a 3D point cloud was created. In the workspace, 10 individual plant species were placed. They could be detected automatically with an accuracy of 2.7 cm. An attached valve was able to spray these specific plant positions, which resulted in a liquid saving of 72%, compared to a conventional spraying method, covering the whole crop row area. 
As plants are dynamic objects, the third objective of describing the plant growth with adequate sensor data, was important to characterise the unstructured agriculture domain. For revering and testing algorithms to the same data, maize rows were planted in a greenhouse. The exact positions of all plants were measured with a total station. Then a robot vehicle was guided through the crop rows and the data of attached sensors were recorded. With the help of the total station, it was possible to track down the vehicle position and to refer all data to the same coordinate frame. The data recording was performed over 7 times over a period of 6 weeks. This created datasets could afterwards be used to assess different algorithms and to test them against different growth changes of the plants. It could be shown that a basic RANSAC line following algorithm could not perform correctly under all growth stages without additional filtering. 
The fourth paper used this created datasets to search for single plants with a sensor normally used for obstacle avoidance. One tilted laser scanner was used with the exact robot position to create 3D point clouds, where two different methods for single plant detection were applied. Both methods used the spacing to detect single plants. The second method used the fixed plant spacing and row beginning, to resolve the plant positions iteratively. The first method reached detection rates of 73.7% and a root mean square error of 3.6 cm. The iterative second method reached a detection rate of 100% with an accuracy of 2.6 - 3.0 cm.
For assessing the robustness of the plant detection, an algorithm was used to detect the plant positions in six different growth stages of the given datasets. A graph-cut based algorithm was used, what improved the results for single plant detection. As the algorithm was not sensitive against overlaying and noisy point clouds, a detection rate of 100% was realised, with an accuracy for the estimated height of the plants with 1.55 cm. The stem position was resolved with an accuracy of 2.05 cm.
This thesis showed up different methods of perception for context awareness, which could help to improve the robustness of robots in agriculture. When the objects in the environment are known, it could be possible to react and interact smarter with the environment as it is the case in agricultural robotics. Especially the detection of single plants before the robot reaches them could help to improve the navigation and interaction of agricultural robots. 

KW  - Roboter
KW  - Landwirtschaft
KW  - Sensor
KW  - Lidar
KW  - Kamera
KW  - Umwelt
CY  - Hohenheim
PB  - Kommunikations-, Informations- und Medienzentrum der Universität Hohenheim
AD  - Garbenstr. 15, 70593 Stuttgart
UR  - http://opus.uni-hohenheim.de/volltexte/2019/1567
ER  -