This thesis addresses vehicle counting in parking lots from UAV aerial footage. We developed a modular system supporting video processing thro\-ugh two pipelines: (i) \emph{mosaic-first}, where frames are stitched into a mosaic before detection, and (ii) \emph{frame-first}, where detections are performed per frame and projected onto the mosaic. The framework includes stitching, homography estimation, and post-processing to filter duplicate detections. Experimental results show that both approaches achieve fairly reliable vehicle counts, while offering complementary advantages in terms of stability and robustness across different parking-lot scenarios.
|
