摘要
目前现有的基于图像的车辆检测系统大多数是利用滑动窗口法来确定车辆候选区域.为了提高车辆检测的速度并减少计算量,提出了一种新的基于图论的车辆检测方法.该方法针对每幅图像通过简单线性迭代聚类(SLIC)算法得到含有若干个超像素节点的图像,分析节点间的相互关系最终确定车辆候选区域.在检测阶段,本文把大量不同视角的车辆图片作为正样本进行训练,得到多视角的分类器;基于候选区域的几何信息,选择适当的多视角分类器进行检测.由公共交通分析数据集(KITTI)检测结果表明:与目前最新的、具有相同提取特征和分类器的算法相比,本文的方法具有更好的检测精度,在复杂的背景下也能取得很好的检测结果.
The majority of the existing graph-based vehicle-detection systems make use of sliding-window paradigm for vehicle-candidate regions location.In order to improve the speed of vehicle detection and reduce the computational complexity,a new vehicle detection method based on graph theory is proposed in this paper.The algorithm uses Simple Linear Iterative Clustering(SLIC)algorithm to obtain images with several super-pixel nodes for each image,and analyzes the relationship among the nodes to determine the vehicle candidate region finally.In the detection stage,multi-view detectors are established by training the vehicle images which seen as the positive samples and collected on each distinct view.Based on the geometrical information of the bounding boxes,the suitable viewpoint detectors are selected from the multi-view detectors.The results of the public traffic analysis dataset(KITTI)show that the proposed approach leads to better performances when compared with the current state-of-the-art methods with the same feature extraction and classifier algorithms.Moreover,it can also yield better results under the complex background.
出处
《北京交通大学学报》
CAS
CSCD
北大核心
2017年第5期66-72,共7页
JOURNAL OF BEIJING JIAOTONG UNIVERSITY
基金
国家自然科学基金项目(61301186
61673047)
北京市科委重大研究专项(SX2016-04)~~
关键词
信息处理
车辆检测
车辆候选区域
多视角分类器
information processing
vehicle detection
vehicle candidate location
multi-view classifiers
