中文摘要：

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。

英文摘要：

In order to obtain the precise dynamic track data from cross camera bicycle under the mixed traffic flow intersections, the experimental research was carried out in regard of the reidentification problem in the process of the cross camera bicycle track. The cross camera bicycle re-identification algorithm under the mixed traffic intersections based on the sample sequence grouping similarity measurement was proposed, with the considerations of complex environment of the mixed traffic intersections, illumination variation and the differences of camera view. In terms of the statistic method, the bicycle sample was divided into three parts and then the ratio of split was cumulated. Through extracting the features from the parts of bicycle gallery, corresponding prototype features were obtained by clustering analysis. With the sample sequence replacing the single sample as a probe, the quantitative analysis of samples was carried out based on comparison analysis. The feature of robustness design was analyzed and a more abstract prototype similarity feature was obtained after the similarity measurement of each sample part and its prototype. Similarity of samples was calculated by within-group linkage and no linkage between groups to improve the time complexity of algorithm by grouping sample sequence with the systematic sampling. In order to analyze the performance of the algorithm, BIKE1, the bicycle reidentification dataset was collected. Meanwhile group performance assessment, prototype parameter settings of components and similar algorithm performance comparison were experimentally compared. The results show that higher identification accuracy is gotten by regarding the sample sequence as a probe, especially when sample sequence is divided into two groups. The bicycle sample, divided into three parts, efficiently strengthens the robustness of influence of algorithm on the illumination variation. Compared with other similar algorithms, identification rate of the above algorithm is much higher.