Browsing by Author "Biggs, DR"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Improving counting performance of densely flocked sheep in aerial imagery.
    (Stellenbosch : Stellenbosch University, 2023-11) Biggs, DR; Schreve, K.; Theart, RP; Stellenbosch University. Faculty of Engineering. Dept. of Mechanical and Mechatronic Engineering.
    ENGLISH ABSTRACT: The use of machine learning and computer vision to count sheep in aerial images taken by an unmanned aerial vehicle (UAV) is explored. The aim is to develop sheep counting approaches which address the challenges related to high object densities and low object-to-image-pixel ratios encountered when using detection-based architectures and aerial images. A comprehensive review of state-of-the-art object detection and instance segmentation techniques and existing object counting approaches is presented. A novel dataset is generated and presented which contains aerial images and videos of sheep grazing in pastures, captured at an altitude of 30m. These images and videos contain scenes of non-uniform object (sheep) distributions and high object densities. A total of 13 different object detection and instance segmentation models are considered, and five of these models undergo hyperparameter optimisation and k-fold cross-validation. The five models are compared based on their crossvalidation mean average precision (mAP) scores on the novel dataset. The best-performing object detection model and instance segmentation model are shown to be Cascade R-CNN and hybrid task cascade (HTC), respectively. The non-uniform distribution of objects, high object densities and low objectto-image-pixel ratios in aerial images present challenges seen in both this dissertation and the literature. It was found that high object densities and low object-to-image-pixel ratios adversely affect counting performance. Two novel approaches are proposed that minimise these effects. The first approach, local density threshold shifting (LDTS), focuses on the challenges posed by high object densities. LDTS shifts the classification probability of each detection based on the density of that detection. This approach achieves an mean absolute error (MAE) and mean absolute percentage error (MAPE) of 26.50 sheep and 4.22%, respectively, on an unseen test dataset and reduces the overall counting error by 78.51% compared to the baseline counting approach. This approach has been published in the International Journal of Remote Sensing. The second approach, sub-window inference, focuses on increasing the object-to-image-pixel ratios of small objects taken in aerial images and videos. This approach utilises a novel cropping technique along with additional data augmentation during training to enhance performance. Sub-window inference achieves an MAE and MAPE of 3.21 sheep and 1.27%, respectively, on an unseen test video dataset, which is a 97.40% reduction in error compared to the baseline counting approach. Sub-window inference has been submitted to the journal Computers and Electronics in Agriculture.