在自然环境下精确实时地检测交通标志小目标对自动驾驶和智慧交通有着重要意义，然而现有算法难以平衡速度与精度的问题。基于YOLOv7-tiny算法，提出了一种改进YOLOv7-tiny的交通标志小目标实时检测算法，即YOLO-T算法。采用条件参数化卷积（CondConv）结构，提升了骨干网络的特征提取能力。为增强对小目标的定位准确度并保证检测速度，设计了TinyFPN特征融合网络结构和ELAN-P网络聚合层。为了验证YOLO-T算法的有效性，在TT100K数据集上做了消融实验和对比实验。实验结果表明，在训练样本及训练设备参数相同的情况下，YOLO-T比YOLOv7-tiny算法的均值平均精度（mAP）提升了16.8%，并且单张图片的检测时间仅10.2 ms。可见，所提的YOLO-T算法能够平衡交通标志小目标的检测速度与精度。

Detecting traffic sign small targets accurately and in real time in natural environment is of great significance for automatic driving and intelligent transportation, however, the existing algorithms are difficult to balance the speed and accuracy. Based on the YOLOv7-tiny algorithm, a real-time traffic sign small target detection algorithm that improves YOLOv7-tiny was proposed, namely YOLO-T algorithm. CondConv (conditional parameterized convolution) structure was used to enhance the feature extraction capability of the backbone network. To enhance the localization accuracy of small targets and ensure the detection speed, the TinyFPN feature fusion network structure and ELAN-P network aggregation layer were designed. To verify the effectiveness of the YOLO-T algorithm, ablation experiments and comparison experiments were done on the TT100K dataset. The experimental results show that YOLO-T improves mAP (mean average precision) by 16.8% over the YOLOv7-tiny algorithm with the same training samples and training device parameters, and the detection time of a single image is only 10.2 ms. Therefore, the proposed YOLO-T algorithm is able to balance the speed and accuracy of the detection of the small targets of traffic signs.