부산대학교 도서관

In recent years, there has been a surge of interest in palm tree detection using unmanned aerial vehicle (UAV) images, with implications for sustainability, productivity, and profitability. Similar to other object detection problems in the field of computer vision, palm tree detection typically involves classifying palm trees from non-palm tree objects or background and localising every palm tree instance in an image. Palm tree detection in large-scale high-resolution UAV images is challenging due to the large number of pixels that need to be visited by the object detector, which is computationally costly. In this thesis, we design a novel hybrid approach based on multimodal particle swarm optimisation (MPSO) algorithm that can speed up the localisation process whilst maintaining optimal accuracy for palm tree detection in UAV images. The proposed method uses a feature-extraction-based classifier as the MPSO's objective function to seek multiple positions and scales in an image that maximise the detection score. The feature-extraction-based classifier was carefully selected through empirical study and was proven seven times faster than the state-of-the-art convolutional neural network (CNN) with comparable accuracy. The research goes on with the development of a new k-d tree-structured MPSO algorithm, which is called KDT-SPSO that significantly speeds up MPSO's nearest neighbour search by only exploring the subspaces that most likely contain the query point's neighbours. KDT-SPSO was demonstrated effective in solving multimodal benchmark functions and outperformed other competitors when applied on UAV images. Finally, we devise a new approach that utilises a 3D digital surface model (DSM) to generate high confidence proposals for KDT-SPSO and existing region-based CNN (R-CNN) for palm tree detection. The use of DSM as prior information about the number and location of palm trees reduces the search space within images and decreases overall computation time. Our hybrid approach can be executed in non-specialised hardware without long training hours, achieving similar accuracy as the state-of-the-art R-CNN.