Unsupervised vehicle extraction of bounding boxes in UAV images

Various studies have been conducted to detect objects in urban areas by applying machine learning algorithms to UAV high-resolution images. However, most vehicle detection studies have limitations in that vehicle detection is performed as a bounding box instead of instance segmentation. Since instance segmentation requires labor-intensive labeling work of each object to train individual objects, research on how to perform unsupervised automatic instance segmentation is needed. Therefore, this study proposed unsupervised SVM classification of the vehicle bounding boxes in UAV images for instance segmentation. As a result of the extraction, it was confirmed that the vehicle could be detected with an accuracy of 89%. It was also confirmed that the vehicle could be detected even if the spectral characteristics within the vehicle were significantly different.