一种面向道路监控视频的车辆智能识别方法

A Vehicle Intelligent Identification Method for Road Monitoring Video

道路监控视频具有场景复杂、车辆型号多样、外观相似、刚性运动等特点.针对道路监控视频中智能识别车辆的应用需求,构建了车辆图像的训练和测试样本库.然后利用Haar-like特征、颜色特征和HOG特征,构建了一个包含车辆形状、边缘、纹理、颜色和梯度等混合特征的细粒度特征池,得到车辆的全局外观模型.在此基础上,构建出一个过完备的初始字典,再通过有监督的判别字典学习方法,训练出一个具有稀疏性和判别性的分类器.为进一步提高视频中车辆判断和识别的稳定性,满足实时视频处理的需求,提出了引入反馈结构约束的"检测—分类—跟踪"车辆识别框架,以提高系统处理视频的实时性和稳定性.

Road monitoring video has the characteristics of complex scenes, diverse vehicle models similar appearance, rigid motion and so on. To meet the application requirements of intelligent identification vehicles in road monitoring video, the training and testing sample library of vehicle images are collected. Then, with Haar-like features, color features and HOG features, a fine-grained feature pool is established, where the shape, edge, texture, color, gradient and other composite characteristics are included, so that an overall appearance model of the vehicle can be deducted. On the above basis, an over-complete initial dictionary is developed. And a classifier with sparsity and discriminability is trained through supervised discriminative dictionary learning method. A vehicle identification framework based on “detection-classification-tracking” and constrained by feedback structure is proposed so that the system’s instantaneity and stability of video processing can be enhanced.