混合交通交叉路口跨摄像机自行车再识别研究

发布时间：2018-11-09 18:08

【摘要】：为了从跨摄像机自行车协同跟踪过程中获得高精度的动态运行轨迹,针对跨摄像机自行车跟踪过程中再识别问题进行了试验研究。考虑到混合交叉路口环境复杂、室外光照变换强烈、摄像机拍摄视角差异等因素的影响,提出了一种在无视域重叠条件下基于样本序列分组相似性度量的混合交通交叉路口跨摄像机自行车再识别算法。采用统计学方法,将自行车样本划分为3个部件并统计出各部件分割比,通过对自行车图库中各部件提取的特征进行聚类分析得到对应部件原型特征。采用对比分析法,将样本序列代替单个样本作为查询图像并对样本数进行了定量分析。从特征鲁棒性设计方面进行分析,通过将每个样本部件与对应部件原型进行相似性度量形成更具抽象性的原型相似度特征。通过系统抽样的方法将图像序列进行分组,并采用组内全连接而组间不连接的方式计算样本间相似度来改善算法时间复杂度。为了有效分析该算法的性能,制作了1个自行车再识别数据集BIKE1,并且在分组性能评估、部件原型参数设置以及同类算法性能比较3个方面进行了试验对比。研究结果表明:采用样本序列作为查询图像具有更高的识别准确率,并且将样本序列分为2组时识别率最高;自行车样本划分为3个部件有效地增强了算法对光照变化等影响因素的鲁棒性;与当前同类算法相比,所提算法具有更高的识别率。
[Abstract]:In order to obtain high-precision dynamic track from the cross-camera bicycle tracking process, the problem of re-recognition in the cross-camera bicycle tracking process is studied experimentally. Taking into account the complex environment of mixed intersections, the intense change of outdoor lighting, and the differences in camera viewing angles, A hybrid traffic intersection cross-camera bicycle rerecognition algorithm based on similarity measurement of sample sequence grouping is proposed in this paper. The bicycle sample is divided into three parts by using statistical method and the partition ratio of each component is calculated. The prototype features of the corresponding parts are obtained by clustering analysis of the features extracted from each component in the bicycle map library. In this paper, sample sequence is used instead of single sample as query image and quantitative analysis of sample number is carried out. Based on the analysis of feature robustness design, a more abstract prototype similarity feature is formed by measuring the similarity between each sample component and its corresponding prototype. The image sequences are grouped by systematic sampling and the time complexity of the algorithm is improved by computing the similarity between samples by using the method of full join within groups and disconnection between groups. In order to effectively analyze the performance of the algorithm, a bicycle recognition data set (BIKE1,) is made and compared in three aspects: group performance evaluation, component prototype parameter setting and performance comparison of the same algorithm. The results show that the sample sequence as the query image has higher recognition accuracy, and the recognition rate is the highest when the sample sequence is divided into two groups. The bicycle sample is divided into three parts which effectively enhance the robustness of the algorithm to the influence factors such as illumination variation. Compared with the current similar algorithms the proposed algorithm has a higher recognition rate.
【作者单位】：华南理工大学土木交通学院;深圳信息职业技术学院交通与环境学院;长沙理工大学综合交通运输大数据智能处理湖南省重点实验室;
【基金】：国家自然科学基金项目(51408237) 湖南省教育厅科研项目(13B132)
【分类号】：U491

【相似文献】