Deep neural networks for object detection in agricultural robotics

摘要

Robotization of tasks in the agricultural domain has the potential to transform food production through continuous surveillance of crops which can facilitate precise admin- istration of nutrients, fertilizers and treatments for weeds and diseases. Such a transfor- mation will increase the sustainability of agricultural practices and improve food security in the future. This thesis applies deep neural network to the task of strawberry detection in video with a view to facilitate surveillance of plant health, crop estimation and logging positions of strawberries. The availability of such data can provide value for growers by enabling op- timization of operations based on observed data, and facilitate progress towards robotic strawberry harvesting. Based on videos sampled from a strawberry growing facility and strawberry images downloaded from the internet, a dataset of strawberries annotated with a state label and coordinates is developed. A set of classi cation models based on deep neural networks are trained on samples from the dataset and applied in a sliding window detection algorithm. Finally uni ed deep neural networks for strawberry detection are trained for the strawberry detection task. Deep neural networks are shown to perform well on the strawberry detection task and real-time processing speeds are demonstrated on an embedded system.