基于HOG纹理的全天时十字路口车尾检测算法  

Study on Algorithm of Vehicle Tail Detection All Day in Crossroads Based on HOG Feature

作　　者：余典[1] 刘操[1] 郑宏[1] 

机构地区：[1]武汉大学电子信息学院,湖北武汉430072

出　　处：《光学与光电技术》2014年第3期18-23,共6页Optics & Optoelectronic Technology

摘　　要：HOG纹理因其良好的鲁棒性,在纹理描述中被广泛使用。提出了一种将HOG纹理应用于十字路口全天候车尾检测的算法。即分别采集了白天和夜间该场景下的车尾作为正样本、非车辆和车辆的一部分作为负样本,经预处理后,提取较低维数的HOG纹理送入支持向量机进行训练,得到白天和夜间的识别模型,在检测中根据一定的条件进行切换。对多段视频进行测试证明,该种算法对不同时段的交通场景都具有较高的稳定的车尾识别率,且优于单模型的识别效果。HOG feature is widely utilized in the description of features for its well robustness. And in this article, an algorithm of applying HOG feature to all-weather vehicle tail detection in crossroads is proposed. Samples in the daytime and in the night-time are collected respectively, while sub-images contain full vehicle tail are extracted from the scenario as positive samples, and non-vehicle sub-images as well as those contain only vehicle parts are taken as negative samples. The low-di- mension HOG features are calculated in the positive samples and negative ones, and then are classified by supportive vector machine based on their sample labels after the pre-processing. In consequence, the recognition models are required, for the daytime and for the night-time respectively. The model will be shifted in detection phase in terms of certain conditions. Sev- eral video tests prove that this algorithm shows relatively high stability and accuracy in vehicle tail detection in different time interval, in comparison with the one using single recognition model.

关 键 词：HOG纹理 支持向量机 十字路口场景 全天时监测 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]