论文部分内容阅读
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。
In order to acquire the high-precision dynamic running track from the cross-camera bicycle collaborative tracking process, the re-identification problem in the cross-camera bicycle tracking process was studied. Considering the complex environment of hybrid intersection, strong outdoor illumination transformation and differences in camera shooting angles, a cross-camera cross-camera bicycle reclassification algorithm based on sample sequence grouping similarity measurement is proposed. . The statistical method was used to divide the bicycle sample into three components and calculate the division ratio of each component. The prototype features of corresponding components were obtained by cluster analysis of the features extracted from each component in the bicycle library. Using comparative analysis, the sample sequence instead of a single sample as a query image and the number of samples were quantitatively analyzed. From the aspects of feature robust design, the more abstract features of prototype similarity are formed by measuring the similarity between each sample component and the corresponding component prototype. The image sequences are grouped by systematic sampling method, and the similarity between samples is calculated by the way of full group connection and no connection between groups to improve the time complexity of the algorithm. In order to effectively analyze the performance of this algorithm, a BIKE1 bicycle recognition data set was made and tested in three aspects: packet performance evaluation, component prototype parameter setting and the performance comparison of similar algorithms. The results show that using the sample sequence as the query image has a higher recognition accuracy, and the recognition rate is the highest when the sample sequence is divided into two groups. Dividing the bicycle sample into three components effectively enhances the influence of the algorithm on the illumination changes Robustness. Compared with the current algorithms, the proposed algorithm has a higher recognition rate.