Automated Analysis of Parking Area Using Drone Imagery

Abstract

Unmanned Aerial Vehicle (UAV) imagery is becoming an influential method for large scale vehicle monitoring, thus enabling smart transportation, urban planning, and parking analytics. Still, the problem of vehicle detection in ultra-high-resolution aerial images persists due to various factors like extreme object downscaling, environmental clutter, occlusions, illumination variation, camera-tilt changes, and the high computational load of deep learning inference. Besides, ground-based and sensor-driven systems have disadvantages like deployment cost, limited viewpoints, and low spatial coverage, which, therefore, necessitates a scalable and accurate aerial detection pipeline. This dissertation presents a refined multi-stage detection strategy for incredibly detailed UAV images. The system merges four main elements: (1) tiled inference with a regulated overlap for small-object continuity, (2) a Soft-NMS-like hybrid suppression method for dense parking scenes, (3) multi-scale inference for the recovery of downsampled objects, and (4) aerial-specific anchor box and augmentation strategies. By employing Faster R-CNN with FPN and efficient context-preserving modules, the pipeline is still achievable on mid-range GPUs. Verification by overlays and ablation studies demonstrates the changes in small-object retention, reduction of boundary loss, and strong performance in crowded parking layouts.