基于改进Yolov8的交通标志检测算法

Improving Yolov8 for Traffic Sign Detection Algorithm

摘要: 针对在自动驾驶行驶中交通标志背景复杂和检测目标小等问题，设计了一种基于yolov8改进的交通标志识别算法。首先在yolov8的DarkNet53中使用动态蛇形卷积，通过引入动态蛇形卷积自适应调整卷积核形状和大小，可以更好地适应小目标尺寸，从而提高准确性，而且可通过目标形状的和轮廓，局部调整卷积核，使在卷积过程中更加关注目标特征，计算复杂度降低，提高检测精度；其次引入局部敏感卷积(LSKA)对SPPF进行改进，通过金字塔池化和频率金字塔来获得多尺度和多频率的特征表示，再利用LSKA来提取局部位置不变性，让其具有更好的鲁棒性和精度，提高网络对交通标志识别的准确性。实验结果表明，改进后的算法在精确率、召回率、平均精确率上相比原算法分别提升2.8%、2.7%、3.3%，检测速度满足实时性要求。

Abstract: A traffic sign recognition algorithm based on Yolov8 improvement is designed to address issues such as complex backgrounds of traffic signs and small target detection in autonomous driving. Firstly, dynamic snake convolution is introduced into DarkNet53 of Yolov8. By dynamically adjusting the shape and size of convolution kernels, adaptive adjustments are made to better accommodate small target sizes, thereby improving accuracy. Moreover, by adjusting convolution kernels locally based on target shapes and contours, attention to target features during the convolution process is enhanced, reducing computational complexity and improving detection accuracy. Secondly, Local Sensitivity Kernel Aggregation (LSKA) is introduced to improve the Spatial Pyramid Pooling Fusion (SPPF). By utilizing pyramid pooling and frequency pyramids to obtain multi-scale and multi-frequency feature representations, and then using LSKA to extract local position invariance, the algorithm exhibits better robustness and accuracy, enhancing the network’s accuracy in traffic sign recognition. Experimental results show that the improved algorithm achieves a 2.8% increase in precision, a 2.7% increase in recall, and a 3.3% increase in average precision compared to the original algorithm, while meeting the requirements for practical detection speed.