目的在行人再识别中,行人朝向变化会导致表观变化,进而导致关联错误。现有方法通过朝向表示学习和基于朝向的损失函数来改善这一问题。然而,大多数朝向表示学习方法主要以嵌入朝向标签为主,并没有显式的向模型传达行人姿态的空间结构,从而减弱了模型对朝向的感知能力。此外,基于朝向的损失函数通常对相同身份的行人进行朝向聚类,忽略了由表观相似且朝向相同的负样本造成的错误关联的问题。方法为了应对这些挑战,提出了面向行人再识别的朝向感知特征学习。首先,提出了基于人体姿态的朝向特征学习,它能够显式地捕捉人体姿态的空间结构。其次,提出的朝向自适应的三元组损失主动增大表观相似且相同朝向行人之间的间隔,进而将它们分离。结果本文方法在大规模的行人再识别公开数据集MSMT17(multi-scene multi-time person ReID dataset)、Market1501等上进行测试。其中,在MSMT17数据集上,相比于性能第2的UniHCP(unified model for human-centric perceptions)模型,Rank1和mAP值分别提高了1.7%和1.3%;同时,在MSMT17数据集上的消融实验结果证明本文提出的算法有效改善了行人再识别的关联效果。结论本文方法能够有效处理上述挑战导致的行人再识别系统中关联效果变差的问题。
This paper introduces the framework of a remote surveillance system and key technologies to implementthe framework. A codec based on MPEG-4 is developed to encode and decode video. Both ratio and speed of compression ...
详细信息
This paper introduces the framework of a remote surveillance system and key technologies to implementthe framework. A codec based on MPEG-4 is developed to encode and decode video. Both ratio and speed of compression of the codec are improved through the optimal designs of motion estimation,quatization and rate control. Errorcorrection functions added to entropy coding of the codec minimize the loss of quality of images due to transmission error. With a QoS mechanism of transmission and playback based on a simple RTP protocol,this framework is able to adapt bandwidth and to control congestion.
暂无评论