Motion blur of vehicle license plates is often caused by high-speed vehicle movement and camera vibration in road traffic surveillance scenarios. This problem results in low localization accuracy and limited adaptability to complex scenes. To improve the accuracy and robustness of license plate detection under motion-blur conditions, an improved YOLOv8n-based license plate localization algorithm, named YOLOv8-KG, was proposed. First, Adaptive Kernel Convolution (AKConv) was used to reconstruct the backbone network and improve feature extraction for blurred license plates. Second, the V-GSCSP structure was introduced into the neck network to reduce computational complexity and model parameters. Then, the Convolutional Block Attention Module (CBAM) was employed to enhance key license plate features. Finally, the Generalized Intersection over Union (GIoU) loss function was adopted to improve bounding-box regression. The results show that YOLOv8-KG achieves an accuracy of 92.9%, a recall of 75.1%, and an average precision of 86.0%. Compared with YOLOv8n, these metrics are improved by 4.3%, 5.7%, and 7.9%, respectively. The number of parameters and computational complexity are reduced by 11.9% and 15.0%, respectively. The results indicate that YOLOv8-KG improves the localization performance of motion-blurred license plates while maintaining a lightweight model. The proposed method provides technical support for real-time license plate detection in complex road traffic surveillance scenarios.