【作者】 赵英男；

【导师】 杨静宇；

【作者基本信息】 南京理工大学 ， 模式识别与智能系统， 2004， 博士

【摘要】 车辆检测与车型识别是智能交通系统(ITS)中的重要组成部分。有鉴于Gabor滤波器在模式识别领域的成功应用，我们将其引入到车辆检测和车型识别应用中，并做了相应的研究工作。包括基于特征加权的Gabor特征抽取算法：基于Gabor滤波器和SVM的红外车辆检测；一种简单的基于Gabor滤波器和边缘特征的车型识别算法：以及实用的基于特定方向的Gabor滤波器组参数设置方法。 文中首先研究并给出了一种改进的基于特征加权的Gabor特征抽取算法。该算法对Gabor特征矢量根据其邻近分量的离散程度进行加权处理，有效增强离散程度相对较小的特征分量在分类中的作用，同时充分利用样本图像的统计信息，具有较强的鲁棒性和类别表征能力。实验数据表明，与传统方法相比，这种特征抽取算法能够有效降低图像识别的错误率，增强鲁棒性，适于对质量较差的图像进行识别。 进一步的，文中将上述算法应用于红外车辆检测中。首先应用阈值分割并结合边检测确定候选区域：其次利用上述Gabor特征抽取算法对选定的车辆和背景样本集进行特征提取，训练SVM分类器；最后应用SVM进行分类检测。实验数据表明，与同类方法相比，该方法在识别率和鲁棒性方面都有所增强。 针对Gabor滤波器应用中的瓶颈问题，即Gabor特征矢量维数较高，以及由此产生的较大计算量和存储负担，文中提出一种简单的基于边缘特征的车型识别算法。不同于目前Gabor滤波器应用中普遍采用的以降低识别率为代价的均匀采样方法，该方法依据车辆具有的明显几何特征，在样本图像的关键部位进行密集采样，非关键部位进行稀疏采样。实验数据表明，这种方法实现简单，在不降低识别率的情况下，有效降低了Gabor特征矢量的维数。 针对目前Gabor滤波器组参数设置算法中存在的不足，即实验法确定的参数不精确，而优化法确定参数又过于复杂，我们提出一种基于特定方向的Gabor滤波器组参数设置方法。该方法根据Gabor特征具有的良好方向特性，首先确定方向参数，然后在每个特定方向进行最佳单Gabor滤波器的参数搜索。我们认为这样得到的Gabor滤波器组，在性能上是接近最优的，同时具有算法简单、数据相关的优点。实验数据表明该算法是实用的、可行的。更多还原

【Abstract】 Robust and reliable vehicle detection and vehicle classification are important issues with applications to Intelligent Transportation Systems (ITS). And we introduce Gabor filters here, since they have been successfully applied for pattern recognition. In this paper we focus our attention on the works in this field, which include four parts: an improved Gabor feature extraction algorithm based on feature weighting, infrared vehicle detection with Gabor filters and Support Vector Machines (SVM), vehicle classification based on Gabor filters and edge features and a practical design for parameters of Gabor filters based on a certain orientation.First, an improved Gabor feature extraction algorithm based on feature weighting is proposed. It weights the raw features derived from 2D Gabor filters according to their own degree of dispersion, which can enhance the effect of the features whose degree of dispersion is relatively small but also can widely used the statistical information of sample images. The experiment results indicate that the proposed method is superior to conventional ones in terms of robustness and discrimination ability. Thus it is fit for the recognition of images with poor quality.Second, an infrared vehicle detection method is developed which contains two main steps: driven hypothesis generation and hypothesis verification. In the hypothesis generation step, possible image locations where vehicles might be present are hypothesized by pixel-dependent threshold selection and edge detection. Hypothesis verification verifies those hypothesis using Gabor filters for feature extraction and SVM for classification. The feature weighting technique mentioned above is used here. This method was tested under four different videos does show visible improvements both in diminishing error rate and robustness.Third, we put forward a novel non-even sampling of Gabor features for classification on the basis of the edge features in vehicles to avoid the heavy computation and memory requirements caused by Gabor feature vectors. In stead of extracting Gabor features at sub-sampled positions of rectangular grid points in general applications, we adopt different sampling intervals on key points and assistant points according to the geometrical features in vehicles. The experimental data show that the method proposed here is simple and effective for both dimension reduction and image representation.Finally, a practical design for parameters of Gabor filters based on a certain orientation is presented. Experiment-based and optimization-based methods are two popular ways in this domain. However, the parameters are not precise in the former and the algorithm is too complex in the latter. To avoid these defects, a more practical one is given. Our main idea is to set the orientation parameters manually based on directional characters in Gabor features, then at each orientation search the optimal single Gabor filter. The parameters of Gabor filters we get are close to optimization, and the algorithm is simple and data dependent as well. The experimental data show that this method is available and efficient.更多还原