Abstract
Agriculture is in a state of constant change. The Austrian trend towards organically produced food, labour shortages and the future wave of farmer retirements are affecting the sector. In this context, automatic systems and robotics are becoming the focus of developments, especially in labour-intensive crops such as vineyard and orchard cultivation.Existing systems are reaching their limits due to dense vegetation, changing light conditions and human activity. The aim of this work is to develop a reliable and cost-effective 3D LiDAR-based detection system. The most suitable algorithms are then
analyzed for different hazardous situations. These scenarios are identified and tested based on a safety analysis.
The procedure includes a comprehensive analysis of the state of the art in outdoor sensor technology and human and object detection using point clouds. The system was built and implemented on the robot “Dionysos”. In practice, this is done using the RANSAC-Algorithm for background removal and an adaptive DBSCAN-Algorithm for clustering.
The two approaches used are a voxel-based approach with Support Vector Machines (SVMs) and a Convolutional Neural Network (CNNs) with 2D depth images.
The integration of new features and a diverse database showed the advantages of SVM, with accuracy exceeding 94 %. The lower computational effort of this approach is particularly noteworthy. With a complete computation time of 0.1671 seconds, this enables a target speed of 1 m/s.
Planned improvements include the extension of the data variety to other crops and hazardous situations as well as the integration of a module for continuous tracking.
These results demonstrate that the use of low-cost 3D LiDAR and computational units in combination with SVMs is a promising solution for human detection in agriculture.
| Date of Award | Oct 2024 |
|---|---|
| Original language | German (Austria) |
| Supervisor | Gerald Zauner (Supervisor) |