<![CDATA[The performance of deep learning-based vehicle detection models often deteriorates when applied to real-world CCTV environments due to domain shift caused by variations in lighting, occlusion, glare, and changes in camera viewpoint. This study aims to develop a YOLOv11-based progressive training framework to improve the model’s generalization ability under heterogeneous operational conditions. The proposed method consists of four stages: base training, fine-tuning, and a supervised progressive domain adaptation strategy implemented through sequential fine-tuning on labeled target-domain CCTV images before OpenVINO-based inference optimization. The model is trained using a source dataset and adapted to a target dataset representing real-world CCTV conditions. Evaluation was conducted using Precision, Recall, mAP@50, mAP@50–95, loss curve analysis, per-vehicle-class evaluation, visual testing on CCTV video, as well as latency and throughput measurements. The results show that base training achieved an mAP@50 of 0.919 and built a robust feature representation, while fine-tuning maintained performance stability with an mAP@50 of 0.915. Although domain adaptation reduced mAP@50 to 0.811 due to domain shift, the model demonstrated improved generalization capabilities and maintained more consistent detection under low-light conditions, glare, occlusion, and heavy traffic. OpenVINO INT8 optimization increased inference speed from 4.86 FPS to 9.08 FPS with minimal accuracy loss. These findings demonstrate that the progressive training framework effectively bridges differences in data distribution while producing a vehicle detection model that is more robust, efficient, and suitable for deployment in real-time CCTV-based traffic monitoring systems.]]>
