计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (4): 147-155.DOI: 10.3778/j.issn.1002-8331.2108-0389

• 模式识别与人工智能 • 上一篇    下一篇

改进SSD算法在交通标志检测中的应用

孙超,温蜜,景俐娜   

  1. 上海电力大学 计算机科学与技术学院,上海 200090
  • 出版日期:2023-02-15 发布日期:2023-02-15

Application of Improved SSD Algorithm in Traffic Sign Detection

SUN Chao, WEN Mi, JING Lina   

  1. College of Computer Science and Technology, Shanghai University of Electric Power, Shanghai 200090, China
  • Online:2023-02-15 Published:2023-02-15

摘要: 针对交通标志检测存在误检率高、鲁棒性差等问题,提出了一种改进SSD(single shot multibox detector)的交通标志检测方法。首先从不同维度提取交通标志的位置和方向感知信息,改善目标在浅层特征图上的感受野区域。其次使用特例化的卷积内核对深层特征图进行条件参数卷积,增强交通标志的特征表达能力。最后对通道注意力机制进行改进,在特征通道中融入目标空间信息,提升交通标志目标的显著性。实验结果表明,提出的方法相较于原始SSD在CCTSDB数据集上的检测精度提升了7.6个百分点,检测速度达到87.5?FPS;在LISA数据集上的平均准确率为94.6%,检测速率为85.0?FPS。相比于其他的检测方法,改进后的SSD算法在复杂的自然场景中对交通标志具有更好的鲁棒性。

关键词: 交通标志检测, SSD, 特征提取, 条件参数卷积, 注意力机制

Abstract: Aiming at the problems of high false detection rate and poor robustness in traffic sign detection, an improved SSD(single shot multibox detector) traffic sign detection algorithm is proposed. Firstly, the location and direction-aware information of the traffic signs are extracted from different dimensions to improve the perceptual field area of the target on the shallow feature map. Secondly, conditional parameter convolution is performed on the deep feature map using a special case convolution kernel to enhance the feature representation of the traffic signs. Finally, the channel attention mechanism is improved to incorporate the target spatial information in the feature channel to enhance the saliency of the traffic sign targets. The experimental results show that compared with the original SSD on the CCTSDB dataset, the detection accuracy of the proposed method is increased by 7.6 percentage points, and the detection speed reaches 87.5 FPS; The average accuracy on the LISA dataset is 94.6% and the detection rate is 85.0 FPS. Compared with other detection algorithms, the improved SSD algorithm has better robustness to traffic signs in complex natural scenes.

Key words: traffic sign detection, SSD, feature extraction, conditional parameter convolution, attention mechanism