Anomaly Detection in Vineyards using Autoencoder

Abstract

In the evolving landscape of agricultural technology, the deployment of autonomous robotic systems in vineyards marks a significant step towards sustainable and efficient farming practices. Among the countless challenges these robotic systems face, anomaly detection emerges as a paramount concern, with implications for both operational efficiency and safety. This thesis delves into the application of Autoencoders, a deep learning technique, for the detection of anomalies within vineyard environments, leveraging the capabilities of these neural networks to identify deviations from normal patterns through unsupervised learning. This study is rooted in the recognition of the complex and dynamic nature of vineyard terrains, which are filled with potential anomalies ranging from natural obstructions to unexpected human or animal presences. The development of a robust anomaly detection system is not merely an academic exercise but a practical necessary for the usage of agricultural robotics. The methodology adopted in this research encompasses the design and implementation of a custom Autoencoder model. The thesis provides a detailed account of the experimental setup, including the configuration of the DIONYSOS prototype robot equipped with advanced sensory technologies, and the systematic approach to model training and validation. The study not only explores the theoretical undertaking of Autoencoders but also extends these concepts to practical applications, demonstrating the adaptability and efficacy of this deep learning technique in real-world scenarios. The findings of this thesis underscore the potential of Autoencoders to revolutionize anomaly detection in vineyard robotics. The results reveal not only the model’s capability to accurately identify a wide range of anomalies but also its efficiency in processing and analysing data in acceptable time, a critical requirement for autonomous robotic operations. Furthermore, the research puts light on the challenges and considerations involved in deploying such technologies in outdoor environments, offering valuable insights for future advancements in the field.

Datum der Bewilligung	2024
Originalsprache	Englisch
Betreuer/-in	Gerald Zauner (Betreuer*in) & Raimund Edlinger (Betreuer*in)