Transfer-Learned Object Detection for Aerial Vehicle Counting: A Small-Dataset Case Study in Drone-Based Parking Management
Urban parking management increasingly relies on automated occupancy counting, but fixed sensors and ground-level cameras suffer from installation cost and mutual vehicle occlusion, motivating aerial, drone-based alternatives. This study presents a drone-oriented vehicle counting system built around a YOLOv8s object detector, fine-tuned by transfer learning from COCO-pretrained weights on a small, single-class Aerial View Car Detection dataset comprising only 280 training images. The model was evaluated on a 19-image held-out test set, achieving a mean average precision at 50 percent intersection-over-union (mAP50) of 0.993, a stricter mAP50-95 of 0.813, a precision of 0.980, and a recall of 0.982, after 100 training epochs on a single cloud GPU. A standalone Python inference application was developed to process folders of aerial images, drawing bounding boxes around detected vehicles and reporting a total count per image, with per-image CPU inference measured at approximately 614 milliseconds. Qualitative inspection of annotated output confirmed reliable detection across varied vehicle colours, parking orientations, and partial edge-of-frame vehicles, with non-maximum suppression correctly resolving overlapping detections in densely packed rows. This study positions its central methodological contribution as an empirical demonstration that transfer learning from a general-purpose, 80-class pretrained detector permits strong single-class aerial detection performance from a training set roughly two orders of magnitude smaller than the datasets typically used to train aerial vehicle detectors from randomly initialised weights, a finding evaluated directly against comparable published work using the same public dataset and against drone-specific parking occupancy systems reported elsewhere in the literature. A genuine inconsistency in the underlying project documentation regarding total training duration is reported transparently rather than silently resolved.
Authors
- Izabo Abigo
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-16
- DOI
- https://doi.org/10.5281/zenodo.24271696
- Primary Topic
- Advanced Neural Network Applications
- Type
- article
- Field-Weighted Citation Impact
- 0.00