交通场景图像中车辆检测和分类研究

发布时间：2018-04-30 19:17

本文选题：车辆检测 + 隐藏变量部件模型　；参考：《北京交通大学》2017年硕士论文

【摘要】：汽车保有量的逐年增加,摄像头的大量应用,使得交通场景中车辆的自动化管理已经成为一大难题。交通场景图像中车辆检测和分类技术是解决这一问题的重要手段,论文选题具有重要的理论意义和实际应用价值。论文主要工作如下:1.给出了一种针对车辆的隐藏变量部件模型训练方法。基于隐藏变量支持向量机,对每类车型都分别训练了隐藏变量部件模型用于车辆检测,模型包含三个部分:主模型、部件模型及部件空间位置关系。车辆模型不仅可以从整体上描述车辆的外观轮廓信息,还可以从细节上描述车辆的部件轮廓信息。实验表明,训练得到的各类车辆模型可以有效的在交通场景图像中检测出车辆的位置。2.给出了一种基于隐藏变量部件模型的车辆分类方法。用训练得到的各类车辆模型分别检测交通场景图像,选择响应值最大模型的检测结果提取车辆图像区域。在提取的车辆图像区域中用所有类别模型进行模型配准,找到最佳的、可以代表各类车型特征的主模型及部件模型位置,能够最大程度的反应车辆的独有信息,具有较大的区分度。提取所有位置的HOG特征作为图像的表示,利用SVM分类器进行分类。经实验表明,同当前已有方法对比,本文所提方法具有更高的分类准确率。3.给出了一种基于卷积神经网络的车辆分类方法。使用卷积神经网络(CNN)对模型配准得到的主模型及部件模型位置进行深度特征提取,将得到的高维深度特征进行主成分分析(PCA),再利用SVM分类器进行分类。实验结果表明,该方法可以有效的提升分类准确率。
[Abstract]:With the increase of vehicle ownership and the application of cameras, the automatic management of vehicles in traffic scene has become a big problem. Vehicle detection and classification technology in traffic scene images is an important means to solve this problem. The topic of this paper has important theoretical significance and practical application value. The main work of this paper is as follows: 1. A training method of hidden variable component model for vehicle is presented. Based on the hidden variable support vector machine (SVM), the hidden variable component model is trained for each type of vehicle for vehicle detection. The model consists of three parts: the main model, the component model and the spatial position relationship of the components. The vehicle model can not only describe the contour information of the vehicle as a whole, but also describe the contour information of the parts of the vehicle in detail. Experiments show that all kinds of vehicle models can effectively detect the position of vehicles in traffic scene images. A vehicle classification method based on hidden variable component model is presented. The traffic scene images are detected by training vehicle models, and the vehicle image regions are extracted by selecting the detection results of the maximum response model. In the extracted vehicle image region, all kinds of models are used for model registration to find the best location of the main model and the component model, which can represent the characteristics of various types of vehicle, and can reflect the unique information of the vehicle to the greatest extent. It has a large degree of differentiation. The HOG feature of all positions is extracted as the representation of the image, and the SVM classifier is used to classify the image. The experimental results show that the proposed method has a higher classification accuracy. 3. A vehicle classification method based on convolution neural network is presented. By using convolution neural network (CNN), the location of the main model and the component model was extracted. The high dimensional depth features were analyzed by principal component analysis (PCA) and then classified by SVM classifier. Experimental results show that this method can effectively improve the classification accuracy.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U495;TP391.41

【引证文献】