Closed networks are IT infrastructures with limited access to the external world, designed to ensure secure data transmission in environments with restricted internet connectivity. These networks offer enhanced protection of sensitive data against external threats. However, most object detection systems in the literature are designed for open networks or cloud environments, and there is limited research on AI-based solutions for closed networks, especially those using thermal imaging. This study presents a customized object detection system designed for closed networks, utilizing images captured by UAVs and thermal cameras and covering various objects such as people, bicycles, cars, and other vehicles. The YOLOv12n model was improved by incorporating GhostConv, BiFPN, the Global Attention Mechanism (GAM), and Fuzzy Sigmoid components. Fuzzy-Based Confidence Scores were applied to address uncertainty in thermal image detection, GhostConv reduced computational costs, GAM focused on important regions, and BiFPN enhanced multi-scale object detection. The customized YOLOv12n model achieved %91 accuracy in the mAP@50 metric, providing a %1.7 improvement over the original model. Additionally, the inference time was reduced from 3 ms to 2.1 ms, making the system %40 faster while maintaining high precision and recall. The trained model enables fast and accurate real-time object detection on-device within secure closed-network environments and was successfully tested on a Jetson platform.