【作者】 李鹏；

【导师】 梁鹏鹏；

【作者基本信息】 郑州大学 ， 工程硕士（专业学位）， 2023， 硕士

【摘要】 多目标跟踪的任务是在视频帧中连续跟踪多个感兴趣的目标,其广泛应用于无人驾驶,安全监控和军事领域等场景。近年来,随着深度学习理论的发展和基准测试的建立,基于深度学习的多目标跟踪方法从表征学习发展到网络建模,通过大量手工标注数据进行监督学习,提高了跟踪的准确性。在夜间复杂场景中,即使是人类也不能清楚的识别物体,获取夜间图像的手工标注是困难和耗时的难题。因此现有的大多数多目标跟踪算法受限于数据集分布不平衡等问题,虽然在光照良好的白天场景下表现出了良好的性能,但是在夜间场景下的跟踪性能下降,制约了多目标跟踪技术的发展。为提高多目标跟踪模型在夜间场景的鲁棒性和泛化能力,本文研究的工作有以下两个方面:(1)针对夜间驾驶场景数据集标注困难的问题,设计了实例可知夜间图像生成算法。首先,使用对称双流网络在多尺度特征上分别解耦出白天图像语义结构和夜间图像实例、背景风格。接着,在空间上对白天图像的实例使用细粒度更高的夜间实例样式,在时序上相同ID对象融合同一种夜间实例样式风格。然后,通过特征自适应去归一化对激活层的输出进行调制,让语义信息在网络中有效地传播并保留内容信息,通过特征自适应实例归一化融合图像的样式风格。最后,在多尺度特征上把有效的内容信息和样式风格进行融合,生成夜间图像。通过在基准数据集上进行白天到夜间风格迁移设计实验,验证了算法的有效性。(2)针对夜间场景多目标跟踪性能下降的问题,设计了可利用自动生成夜间数据的多目标跟踪算法。本文首先利用实例可知夜间图像生成算法生成夜间图像,白天与夜间图像之间的全局场景结构一致,夜间图像可直接使用白天图像的标注信息。然后,生成的夜间图像与对应的标注集合构成了合成数据集。最后,通过增加有标注的夜间数据集,平衡数据集中白天域与夜间域样本的分布,减小域差距并增强模型的鲁棒性。在3种多目标跟踪算法上进行实验,验证了此方法对提升模型在夜间场景的泛化能力是有效的。更多还原

【Abstract】 The task of multiple object tracking is to continuously track multiple targets of interest in video frames,which is widely used in scenarios such as unmanned driving,security monitoring and military fields.In recent years,with the development of deep learning theory and the establishment of benchmark tests,multiple object tracking methods based on deep learning have developed from representation learning to network modeling,and supervised learning through a large amount of manually labeled data has improved tracking accuracy.In complex scenes at night,even humans cannot clearly identify objects,and obtaining manual annotation of nighttime images is a difficult and time-consuming problem.Therefore,most of the existing multiple object tracking algorithms are limited by the unbalanced distribution of data sets.Although they show high performance in daytime scenes with good lighting,the tracking performance in night-time scenes decreases,which restricts multiple object tracking.technology development.In order to improve the robustness and generalization ability of the multiple object tracking model in night scenes,the research work of this paper has the following two aspects:(1)Aiming at the difficulty of labeling the night driving scene dataset,an instance aware night image generation algorithm is designed.First,a symmetric two-stream network is used to decouple the semantic structure of the daytime image and the instance and background style of the nighttime image on multi-scale features.Then,use a finer-grained nighttime instance style for daytime image instances in space,and fuse the same nighttime instance style with the same ID objects in time sequence.Then,the output of the activation layer is modulated by feature-adaptive denormalization,so that semantic information can be effectively propagated in the network and content information is preserved,and the style of the fusion image is fused by feature-adaptive instance normalization.Finally,effective content information and style are fused on multi-scale features to generate nighttime images.The effectiveness of the algorithm is verified by conducting day-to-night style transfer design experiments on benchmark datasets.(2)Aiming at the performance degradation of multiple object tracking in night scenes,a multiple object tracking algorithm that can use automatically generated night data is designed.This paper firstly uses the instance aware that the night image generation algorithm generates night images.The global scene structure between the day and night images is consistent,and the night images can directly use the annotation information of the day images.Then,the generated nighttime images and corresponding annotation sets constitute a synthetic dataset.Finally,by adding labeled nighttime datasets,the distribution of daytime domain and nighttime domain samples in the dataset is balanced to reduce the domain gap and enhance the robustness of the model.Experiments on three multiple object tracking algorithms have verified that this method is effective in improving the generalization ability of the model in night scenes.更多还原