图切技术和卷积网络的交通标志数据集构建方法

Traffic sign dataset constructing method with convolutional neural network and graph cut

为解决自然场景下的交通标志数据获取困难的问题,提出一种基于深度卷积神经网络结合图切技术的交通标志数据自动收集方法。该方法先利用人为收集的少量7大类交通标志数据集训练检测交通标志的卷积神经网络模型,利用该网络模型检测图像或视频中的交通标志类别、位置及可信度,保存大于给定阈值的交通标志信息;然后采用图切技术对检测的交通标志进行分割得到精度更高的交通标志区域,将此标志的区域信息和类别作为标定信息。将对应的图片和标定信息按要求生成新的训练数据集和测试数据集,重新微调训练生成新的网络模型。实验结果表明,重新微调训练的网络比初始网络的平均检测精度提升了6.6%。该方法可将车载相机或是行车记录仪等设备获取的图像或是视频中的交通标志自动保存下来生成构建新的交通标志数据集,省去人工标定的过程。

In order to solve the problem that it is difficult to get traffic signs from natural scenes, an automatic traffic sign collecting method based on deep Convolutional Neural Network（ CNN） and graph cut was proposed. Firstly, an initial dataset was obtained by collecting seven main categories of traffic signs and their subclasses. The initial dataset was used to train CNN model for traffic sign detection. Then the traffic sign＇s category, confidence level and location of images or videos were detected by the trained CNN model. Secondly, the detected traffic sign＇s region greater than a given threshold was stored. The corresponding new region was obtained by localization algorithm based on graph cut. Thirdly, the new traffic sign＇s region and its category were used as its ground truth. The new training set and testing set by adding the new detected traffic signs were applied to fine-tune the convolution network and obtain a new model. The experimental results show that the fine-tuned model has a higher mean average precision than the initial one. This method can automatically generate a new traffic sign data set by relocating the detected traffic signs from images or videos which were captured by the vehicular camera or the traveling data recorder.