<p>Autonomous vehicles require fast and accurate perception under varying environmental conditions, but most high-performance detectors are too heavy for edge devices. A lightweight yet robust model is needed to ensure safety and real-time operation in practical scenarios. This work proposes to develop a lightweight traffic-sign detection system suitable for real-time resource constrained embedded platforms. The objective of this study is to enhance the YOLOv8n architecture through an effective attention mechanism and validate its embedded performance on a real autonomous vehicle platform. To achieve this, a custom traffic-sign dataset was created and used to train an improved YOLOv8n model integrated with the proposed Turbo Convolutional Block Attention Module (TurboCBAM), and its robustness was further evaluated across diverse datasets. The TurboCBAM model achieved an mAP@50 of 0.977, with inference time increasing from 5.5 ms in YOLOv8n to 6.2 ms. During real-time embedded testing, it maintained a stable performance of 10 to 12 frames per second (FPS). To validate its robustness and generalization capability, the model was tested on external datasets such as TT100K and CCTSDB2021. It achieved mAP@50 scores of 0.860 and 0.784, demonstrating reasonable performance under domain shifts, with some degradation due to dataset differences. These cross-dataset results show that the model maintains reliable detection across diverse environments and performs effectively beyond the conditions seen during custom dataset training.</p>