Development of YOLO-based object detection and tracking systems for airport ground safety and simulation-based collision analysis

Abstract

Airport ground incidents remain one of the major safety concerns in aviation, causing operational disruption and substantial economic losses. Although conventional surveillance and sensing can support surface awareness, cost, deployment constraints, and limited interpretability at close range motivate low-cost, vision-based alternatives. Key research gaps include (i) apron-specific, part-level, openly available labelled datasets, (ii) systematic and fair benchmarking of modern deep learning architectures under consistent evaluation conditions, and (iii) integrated risk analysis that progresses from perception to actionable early warning. This thesis addresses these gaps by developing and validating a scalable early-warning framework based on Computer Vision (CV) and Deep Learning (DL), designed for integration with existing Closed-Circuit Television (CCTV) infrastructures. The thesis follows a multi-stage methodology. First, a new detection and segmentation dataset was constructed with five aircraft classes (airplane, wing, nose, tail, and fuselage) to support part-level perception for apron safety. Using this dataset, twelve modern object detection and segmentation architectures were trained and evaluated under consistent experimental settings to establish a benchmarking baseline. YOLOv8-Seg (You Only Look Once, version 8-Segmentation) emerged as the most suitable backbone for the intended operational constraints. Second, to improve robust detection and segmentation of aircraft and their components, YOLOv8-Seg was systematically optimised through a six-step, ablation-driven pipeline, spanning parameter tuning, loss-function refinement, data augmentation, and inference-efficiency improvements. Third, a Multi-Object Tracking (MOT) dataset was created and annotated in the MOTChallenge format to benchmark leading trackers under identical evaluation settings. Finally, two complementary safety layers were developed: (i) a reactive module that issues immediate alerts using image-plane geometric proximity derived from segmentation outputs, and (ii) a proactive module that forecasts short-horizon conflicts by extrapolating past motion and evaluating future overlap using Intersection over Union (IoU). Experimental results show that the optimised YOLOv8-Seg model significantly improves segmentation performance by +8.04 points in mAP@0.5:0.95 (mean Average Precision averaged over IoU thresholds from 0.50 to 0.95) and +4.74 points in mAP@0.5. On the tracking benchmark with airplane-only ground truth, DeepSORT achieved the strongest overall performance, reaching 92.77% Multi-Object Tracking Accuracy (MOTA) with 93.27% recall. The framework was validated across multiple representative scenarios using both simulated and real-world video, supporting the feasibility of a low-cost, CCTV-compatible approach for enhancing apron safety through integrated perception, tracking, and early warning.