Research on black-and-white image processing method of smart car camera

In view of the images collection and processing problems of the smart car camera, the paper introduces a method which deals with field and line synchronization signal separation and binarization processing of the video signal collected from track fields, and which is capable to extract and position black border trajectory images effectively. According to the experiment results of the method, the camera images can be collected and processed effectively, and the accurate image information can be provided for the smart cars to travel along the track. The method has the advantages of being easy to use, strong adaptability, ideal performance and high practical value. On the basis of advantages the method is of high practical value in smart car races.

结合时空距离的多网络互学习行人重识别

目的在真实行人识别场景中,获得准确的标注需要耗费大量人力,因此无监督领域自适应成为行人重识别具有潜力的研究方向,这类方法通常需要聚类生成伪标签,往往会存在噪音。此外,在行人搜索过程中,好的排序算法也是取得更好识别性能的关键,但寻常的Re-Ranking排序优化由于巨大的性能消耗,限制了在真实场景下的应用。针对这两个问题,本文提出了一个联合多网络、分摄像头训练的框架,利用时空信息对排序进行优化。方法对源域数据使用有监督进行预训练,利用未标记的目标域样本进行多个网络模型的深度互学习无监督训练,提高网络的泛化能力,同时在训练过程中进行分摄像头处理,减小跨摄像头的影响,提升伪标签的质量。在排序匹配阶段利用时空信息对排序进行优化,进一步提升匹配性能。结果实验在2个跨域实验数据集上进行测试比较,在源域为DukeMTMC-ReID(Duke multi-tracking multi-camera re-identification)数据集,目标域为Market-1501数据集的实验中,本文方法的平均精度均值(mean average precision,mAP)和Rank1分别为82.5%和95.3%;在源域为Market-1501,目标域为DukeMTMC-ReID数据集的实验中,mAP和Rank1分别为75.3%和90.2%。结论提出的结合时空距离排序的分摄像头网络互学习模型,提升了伪标签的精度,并优化了匹配排序,相比于其他优化算法大幅减少了计算量,进一步提升了行人重识别性能。