【摘要】 针对真实场景中小目标交通标志检测精度低的问题，提出一种改进YOLOv5的小目标交通标志检测方法。对原主干网络进行简化，降低网络的复杂度；使用高分辨率特征融合网络以减少特征融合时分辨率的损失；在保持三尺度检测的前提下引入大尺寸检测头，提升对小目标的检测能力；引入CBAM注意力机制，挖掘有关小目标的特征信息；引入SPD-Conv取代网络中的跨步卷积，提升特征学习的效果。在TT100K数据集上的实验结果表明，所提方法在小目标交通标志上的检测精度为79.3%,相较于原YOLOv5算法提升了6.1%,整体检测效果优于YOLOX等主流目标检测算法，算法的检测速率为39.4 f/s,满足实时检测的需求。更多还原

【Abstract】 Aiming at the low accuracy problem of small target traffic sign detection in real scenes, a method was proposed to improve YOLOv5 for small target traffic sign detection. The original backbone network was simplified, which reduced the complexity of the network. The high-resolution feature fusion network was used to reduce the loss of resolution during feature fusion. On the premise of maintaining three-scale detection, the large-size detection head was introduced to improve the detection ability of small targets. The CBAM attention mechanism was introduced, which mined feature information about small objects. To improve the effect of feature learning, SPD-Conv was introduced to replace the strided convolution in the network. Experimental results on the TT100K dataset show that the detection accuracy of the proposed method on small target traffic signs is 79.3%, which is 6.1% higher than that of the original YOLOv5 algorithm, and its overall detection effect is better than that of mainstream target detection algorithms such as YOLOX. The detection rate of the algorithm is 39.4 f/s, which meets the requirement of real-time detection.更多还原