遮挡状态下行人再识别方法研究

Research on pedestrian re⁃recognition method under occluded state

随着相关计算机视觉算法的迅速发展,视频监控系统大规模的投入使用,不但提高了交通安全,也促进了智慧高速的发展。行人再识别算法作为视频监控系统中的重要技术环节日渐成熟与应用,进一步促进了高速公路建设与运营的安全。然而,由于应用场景的复杂性,尤其是在面对繁杂的场景遮挡因素时,伴随着遮挡物产生的噪声影响,不可避免地造成了被识别人或物的特征信息损失,对现有的行人再识别算法提出了极大的挑战。研究过程中,通过有针对性地遮挡样本生成及提高训练效率的损失函数的构建,提出基于注意力机制的网络结构。通过实验比较分析,提出的算法与近期的讨论热点相比有一定的优势,消融实验分析也进一步证明了所提出的网络结构的有效性。

With the rapid development of relevant computer vision algorithms and the large⁃scale deployment of video surveillance system,the traffic safety has been improved and the development of smart highways has also been promoted.As an important technique of video surveillance system,the pedestrian re⁃identification algorithm is increasingly mature and applied,which further promotes the safety of highway construction and operation.However,due to the complexity of the application scene,especially in the face of complex scene occlusion factors,accompanied by the noise generated by the occlusion objects,it inevitably causes the loss of the feature information of the identified person or object,which poses a great challenge to the existing pedestrian re⁃recognition algorithm.In this research,the network structure based on the attention mechanism is proposed by constructing loss function to generate occlusion samples and improve the training efficiency.The result of experimental comparative analysis proves that the proposed algorithm has certain advantages in comparison with the recent discussion hot⁃spots.The ablation experimental analysis demonstrates the validity of the proposed network structure as well.