With the increasing popularity of self-driving cars, road-condition detection systems have become a significant research focus. Traffic sign detection, which is a crucial component of these systems, directly affects the safety of both drivers and pedestrians. Owing to the urgent requirement for efficient traffic sign detection for autonomous driving applications, achieving high-performance and rapid responses for long-distance sign detection is crucial. You only look once v7 (YOLOv7) is a one-stage object detection model that offers excellent detection speed but faces challenges in long-range detections owing to the inherent loss of small-object features in its convolutional and maxpooling layers. To address these challenges, we propose enhancements for YOLOv7 by integrating a space-to-depth convolution module to better preserve small-object features and an attention mechanism to help it focus more effectively on relevant objects. We further enhanced it by adding extra detection heads specifically designed to extract small-object features and incorporated Gaussian noise to enhance its robustness. The improved model was evaluated on the National Taiwan University of Science and Technology Taiwan traffic sign dataset, which comprises 29 types of traffic sign. The results demonstrated the effectiveness of these enhancements, improving the mAP50 of YOLOv7 from 59.5 to 84.7% and offering a significantly better traffic sign detection performance.