智能行车中车辆检测跟踪与限速标志识别是智能交通系统领域中非常重要研究方向，也是自动驾驶、辅助驾驶等技术的基础，本文主要以单目视觉为基础，在智能行车框架下对车辆检测、车辆跟踪以及交通限速标志识别进行深入研究。 本文研究了基于区域假设与验证的车辆检测方法。在假设区域生成阶段，利用基于BLOB感兴趣区域划分方法，结合对称性生成车辆假设区域。在假设区域验证阶段，本文首先提出了基于结构自相似度特征的验证方法，将结构自相似性指标纳入特征生成框架，然后利用SVM训练分类器进行检测。在此基础上，提出融合两个方向的HOG特征并结合AdaBoost分类器验证假设区域的方法。并通过实验验证了车辆检测方法的有效性。 提出了一种基于混合注意力的车辆检测技术，结合注意力机制对全卷积神经网络进行改造，在全卷积神经网络上附加了通道注意力和空间注意力，使车辆检测效果提高了2.73%，达到了98.97%。 在车辆检测的基础上，论文对车辆跟踪方法进行进一步研究。提出了一种改进的L1-跟踪算法，算法以粒子滤波框架的L1-跟踪算法为基础，基于Haar-like特征的计算方法对传统L1-算法过完备基构造方式升级。使用图像分块的方法进行降维，极大缩小了运算复杂度。而且，算法在不损失精度的前提下减少了模板数量，再次提高了计算效率。 考虑我国目前的环境状况，论文对霾天气下限速标志识别方法进行了研究。在对暗通道法研究的基础上，引入最大全局光照强度参数，解决了暗通道法光晕和失真等问题。提出融合颜色、形状以及纹理多种特征的限速标志检测方法，首先利用颜色和形状在图像中筛选出交通标志候选区域，使用HOG纹理特征结合SVM分类器完成候选区域的验证，最后利用卷积神经网络完成交通标志的识别。 本文分别使用不同的数据库对本文研究的车辆检测、车辆跟踪以及限速标志识别算法进行验证。实验结果说明，与传统经典算法相比，本文提出的方法具有更好的识别率以及时间效率。 关键词：单目视觉；车辆检测；混合注意力；车辆跟踪；交通限速标志识别

 Vehicle detection,tracking and speed limit sign recognition in intelligent driving were important research directions in the field of intelligent transportation systems, were also the basis of autonomous driving and assisted driving technologies. This thesis mainly studied the vehicle detection,vehicle tracking and traffic speed limit sign recognition based on monocular vision under the framework of the intelligent traffic system. Vehicle detection method based on region hypothesis and validation was studied. In the hypothesis region generation stage,BLOB-based ROI partition method was used to generate vehicle hypothesis region combined with symmetry. In the hypothesis region verification stage,this paper proposed a verification method based on structural similarity feature, which incorporates structural similarity index into the feature generation framework, and then use SVM training classifier to detect. On this basis,a method of fusing HOG features of two directions and verifying hypothetical regions with AdaBoost classifier was proposed. Besides, the validity of the vehicle dection method is verified by expermients. A mixed attention-based vehicle detection technology was proposed, which combines the attention mechanism to full convolutional neural network, channel attention and space attention were added to the full convolutional neural network, which improved the vehicle detection effect by 2.73% to 98.97%. On the basis of vehicle detection,the vehicle tracking method was studied. An improved L1-tracking algorithm was proposed. Based on the L1 and particle filter framework. To improve the method of construction of the proposed method the over-complete base,the proposed method reduces the dimensions of feature by blocking image samples, which greatly reduces the computational complexity. At the same time,the number of templates was reduced, which reduces the number of sparse calculations again. Considering the current environmental conditions in China,the recognition method of the traffic speed limit sign in haze weather was studied. Firstly,based on the defogging algorithm based on the dark channel prior principle, the maximum global illumination intensity parameter was introduced in the calculation of the global illumination intensity in the original algorithm,preventing the questions of color spots and distortion. A method of detecting the speed limit sign combining various features of color, shape and texture was proposed. Firstly, the traffic sign candidate area was selected in the image by using color and shape, and the candidate area was verified by using the HOG texture feature combined with the SVM classifier. Finally, the convolution was used. Finally, convolution neural network was used to identify traffic signs. Different databases were used to verify the vehicle detection, vehicle tracking and speed limit sign recognition algorithms studied in this thesis. The experimental results show that compared with the traditional classical algorithm, our methods have better recognition rate and time efficiency. Key words: monocular vision; vehicle detection; mixed attention; vehicle tracking; traffic speed limit sign recognition