基于特征融合与稀疏编码的交通标志识别算法研究

发布时间：2018-10-25 09:00

【摘要】：交通标志识别是汽车辅助驾驶系统和无人驾驶系统的关键问题之一。由于交通标志位于户外,使其极易受到天气、光照等因素的影响。同时,还存在遮挡、视角变化、因汽车运动而造成的图像模糊等问题。因此,大类别集交通标志识别在具有广阔应用前景的同时也具有很大的挑战性。交通标志的预处理、特征提取以及分类器设计是交通标志识别中的关键问题。本文针对大类别集交通标志识别中的特征提取环节进行了深入的研究。论文的研究工作主要包括以下三个方面：(1)方向梯度直方图(HOG)和尺度不变特征转换(SIFT)是目前普遍采用的特征,论文在对HOG特征与SIFT特征研究的基础上,给出了一种改进的融合HOG和SIFT的特征提取方法。该方法提取的特征不仅具有HOG特征因区块直方图统计带来的抗干扰性,而且还具有SIFT特征因主方向匹配带来的旋转不变性。因此,该特征可提高对大类别集交通标志图像的特征表达能力,为标志识别打下基础。(2)为进一步对特征进行优化,本文采用了一种对特征进行二次编码的算法。该算法采用局部稀疏编码(LLC)方式,并针对大类别集交通标志进行了参数优化；在基准空间的选择上,给出了一种将基于样本学习的K-SVD优化基准空间与基于数学模型的离散余弦变换(DCT)基准空间进行级联的多级基准空间。(3)采用上述优化后的特征与支持向量机(SVM)实现交通标志的识别,对58类交通标志,总计2494张图像进行识别,取得了99.20%的分类正确率,仿真实验结果验证了本文方法的有效性。
[Abstract]:Traffic sign recognition is one of the key problems in vehicle aided driving system and driverless system. Traffic signs are easily affected by weather, lighting and other factors because of their outdoor location. At the same time, there are some problems, such as occlusion, angle of view change, image blur caused by automobile motion and so on. Therefore, large class set traffic sign recognition has a wide application prospect and a great challenge at the same time. The preprocessing, feature extraction and classifier design of traffic signs are the key problems in traffic sign recognition. In this paper, the feature extraction in large class set traffic sign recognition is studied. The research work of this paper mainly includes the following three aspects: (1) the directional gradient histogram (HOG) and scale invariant feature conversion (SIFT) are widely used at present. An improved feature extraction method combining HOG and SIFT is presented. The feature extracted by this method not only has the anti-interference property of HOG feature due to block histogram statistics, but also has the rotation invariance caused by SIFT feature due to the main direction matching. Therefore, this feature can improve the ability of feature expression for large class traffic sign images and lay the foundation for sign recognition. (2) in order to further optimize the features, this paper adopts a quadratic coding algorithm for features. The algorithm adopts the method of local sparse coding (LLC), and optimizes the parameters for large class traffic signs. In this paper, a multilevel reference space is proposed, which cascades the K-SVD optimization reference space based on sample learning and the discrete cosine transform (DCT) reference space based on mathematical model. (3) the optimized feature and support vector machine (SVM) are used to realize the optimization. Identification of traffic signs, A total of 2494 images of 58 kinds of traffic signs are recognized and the classification accuracy is 99.20%. The simulation results show that the proposed method is effective.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】