CASIA OpenIR  > 毕业生  > 硕士学位论文
基于平行视觉的交通标志识别方法研究
陈然1,2
学位类型工程硕士
导师王飞跃
2018-05-24
学位授予单位中国科学院大学
学位授予地点北京
关键词平行视觉 交通标志识别 合成样本 域相关 Ls-gan
摘要      交通标志识别是智能交通系统的重要组成部分,一般可划分为检测与分类两个阶段。检测是指从图像或视频序列中提取出可能含有交通标志的目标区域,分类是对该目标区域进行识别,得到交通标志的类别。交通标志识别技术可以用于无人驾驶、辅助驾驶以及交通标志的日常维护,具有广阔的应用前景。传统的识别方法利用交通标志对称的形状特征以及均匀、鲜明的颜色分布来检测和分类交通标志,这些方法取得了一定的成果,但是性能仍然有待进一步提高。伴随着深度学习的兴起,越来越多的研究者将深度学习方法应用于交通标志的识别,并取得了超越传统方法的效果。但是,基于深度学习的方法需要大量且多样化的训练样本,在实际中,标记样本是一件费时费力的事情,而且有些场景下的训练样本并不容易获取。平行视觉方法提供了一种可行的解决方案,该方法提出构建人工场景,并自动生成训练样本以及标注信息。由于人工场景是可调控的,因而可以生成任意状态下的样本并得到准确的标注信息。除此之外,平行视觉方法还强调虚实互动,也就是将视觉模型在实际场景与人工场景中平行执行,使模型训练和评估在线化、长期化。本论文利用深度学习技术,结合平行视觉理论,对自然环境下的交通标志检测与分类进行了研究,主要工作如下:
      首先,对原始数据集,本论文进行了预处理,并由此提出一种保存完整目标区域的图像分割方法,该方法完整保留了图像中的目标,并将大尺寸的图像分割成若干张小尺寸的图像。原始数据集中,训练样本的尺寸太大,而有的交通标志所占的区域太小,不适合直接进行压缩。利用裁剪的方法,去除了大部分背景,但对图像中的目标区域没有影响。
      其次,该论文对合成训练样本进行了探讨和研究。该方法将真实图像中的交通标志裁剪出来,经过一系列几何形变、亮度变化,再与背景图像合成新的训练样本。几何形变、亮度变化能够很好地模拟自然条件下发生的变化,使得样本的多样性得到加强。该论文除了生成合成样本之外,对训练方法也进行了调整,也就是在训练过程中,将真实样本与合成样本一起,共同训练模型。在TT100K和GTSDB数据集上,通过一系列的对比实验,表明了上述方法确实具有有效性。
      再次,研究了利用LS-GAN生成虚拟交通标志,并与真实交通标志一起训练模型。LS-GAN是一种改进版本的GAN,该方法并不是简单地复制原数据集中的样本,而是生成新的、具有多样性的样本。对于无标签的交通标志,提出利用半监督方法中的域相关方法进行分类。该方法能够在有标签的交通标志与无标签的交通标志之间建立联系,从而使得无标签的交通标志即使没有标注信息,也能够得到训练。相比较于只利用有标签的交通标志进行训练,该方法能够大大提高无标签的交通标志的分类准确率。
      最后,构建了交通标志识别系统,能够对视频中的交通标志进行准确识别。该系统首先将视频分成连续的多帧图像;然后利用训练好的模型对每帧图像进行交通标志识别;最后,将识别结果转换为视频并实时显示。
其他摘要      Traffic sign recognition (TSR) is an essential part in intelligent transportation systems (ITS), and it includes traffic sign detection (TSD) and traffic sign classification (TSC). TSD is a process that extracts region proposals from images or videos and TSC is a process that classifies traffic signs. TSR can be used for intelligent autonomous vehicles, driver support systems and sign inventory, so this technique can be used wildly. Traditionally, the symmetric shapes and brilliant colors are utilized for TSR. However, in the chaos of some situation, traffic signs are hard to be detected or classified correctly. With the flourish of deep learning, more and more CNN-based methods are applied for TSR, and get preferable results. CNN-based methods need many labeled samples, but labeling samples is a time-consuming work, and some samples are not easy to obtain. Parallel vision method provides an alternative solution to this problem. This method generates labeled samples by constructing artificial scenarios. Artificial scenarios are controllable, so diversified labeled samples will be generated by this method. Besides, parallel vision proposes to train and evaluate models in both real scenarios and artificial scenarios, which is named parallel execution. This thesis combines deep learning with parallel vision to explore the detection and classification of traffic signs. The main contributions of this paper are described as follows:
      First, this thesis preprocesses the original dataset, and proposes an image segmentation method which preserves the total regions of traffic signs, while crop an original sample into several small samples. In original samples, the object regions are small and not suitable to resize samples. By using cropping, many background is excluded, but object regions are preserved.
      Second, this thesis proposes a method to synthesize samples. In this method, first, traffic signs are cropped from original samples according to annotation, then, a series of variations are applied to traffic signs, including distortion, rotation, light variation, at last, varied signs are pasted to background images to get synthesis samples. In this method, distortion, rotation and light variation imitate the variation in natural scenarios. Besides synthesizing samples, in this thesis, real samples and synthesis samples are used to train model jointly. By a series of contrast experiments on TT100K dataset and GTSDB dataset, the superiority of the proposed method is proven.
      Third, this thesis adopts LS-GAN to generate virtual samples, and combines virtual samples with real samples to train a model. LS-GAN is a modified GAN, and can generate diversified traffic signs. Besides, this thesis uses domain association for classification of unlabeled samples. Domain association is a semi-supervised method for classification, and can establish relationship between labeled traffic signs and unlabeled traffic signs. During training, although some samples are unlabeled, they still make a contribution to model training. In the following test stage, the model trained on both samples get better performance for unlabeled sample classification.
      At last, this thesis codes a network for real-time traffic sign recognition. First, a series of images are extracted from a test video, then, traffic signs in these images are recognized by well-trained model, at last, the recognized results are converted to a video and shown in real-time.
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/21188
专题毕业生_硕士学位论文
作者单位1.中国科学院大学
2.中国科学院自动化研究所
推荐引用方式
GB/T 7714
陈然. 基于平行视觉的交通标志识别方法研究[D]. 北京. 中国科学院大学,2018.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
硕士论文-陈然.pdf(15145KB)学位论文 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[陈然]的文章
百度学术
百度学术中相似的文章
[陈然]的文章
必应学术
必应学术中相似的文章
[陈然]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。