【作者】 李洁；

【导师】 杨明静；

【作者基本信息】 福州大学 ， 电子信息（专业学位）， 2023， 硕士

【摘要】 视觉目标跟踪的目的是在一个视频的不同帧中识别出同一个目标,从而进一步分析目标的行为,该技术已被广泛应用于智能监控系统和智能交通系统。现阶段的视觉目标跟踪技术较少关注跟踪技术在夜间场景下的表现,与正常光照场景相比,夜间目标跟踪不仅需要面对跟踪任务中常见的运动模糊、物体间的遮挡和尺度变化等挑战,还需面对夜间图像中光照条件不足与灯光干扰等问题。为了改善实时视觉目标跟踪算法在夜间场景下的表现与扩展视觉目标跟踪技术的应用范围,本文针对夜间目标跟踪存在的问题,设计出基于低光照图像增强的解决方案,将低光照图像增强模块用于图像预处理阶段,连接实时的跟踪器进行目标跟踪。本文从监督学习和无监督学习的方式出发,提出两种低光照图像增强算法。本文的主要研究成果如下:一、提出基于Retinex理论的低光照图像增强方法RAUNet。针对夜间图像光照条件较差的问题,本文基于Retinex理论提出了一个反射分量估计网络RAUNet,将反射分量作为低光照图像增强的结果。方法基于U-Net结构,引入注意力模块提取更有效的特征,降低噪声的影响。基于深监督思想设计了多尺度监督模块,促进网络训练的快速收敛和生成更高质量的图像。结合实时跟踪方法在公开数据集上的测试结果表明,夜间图像在经过RAUNet增强后再送入跟踪网络的方式可以有效提高跟踪的成功率和精确度。二、提出基于生成对抗网络的低光照图像增强方法RDEnlighten GAN。RAUNet基于监督学习的方式需要同一个场景中低光照和正常光照图像组成的图像对,而此类图像对不易获取。夜间场景中除了光照条件较差之外,灯光的干扰会导致图像中出现光照不均现象,损失部分细节特征。针对上述问题,本文在生成对抗网络的基础上设计了RDEnlighten GAN方法。生成器结合改进的残差密集块,提取更有效的局部特征,使用自正则化的注意力模块,有效保持图像的亮度均匀。采用全局和局部判别器,保证全局和局部区域细节的真实性。联合跟踪方法在夜间跟踪数据集上的实验结果证明了RDEnlighten GAN方法的有效性。综上所述,针对夜间目标跟踪存在的光照不足与光照不均问题,本文提出了基于Retinex理论的低光照图像增强方法和基于生成对抗网络的低光照图像增强网络,结合实时目标跟踪方法在夜间目标跟踪数据集上的测试结果验证了本文方法的有效性。更多还原

【Abstract】 The goal of visual object tracking is to recognize the same target in different frames of a video,in order to further analyze its behavior.This technology has been widely applied in intelligent surveillance systems and intelligent transportation systems.Currently,there is a lack of emphasis on the performance of visual object tracking during nighttime conditions.In comparison to normal-light conditions,visual object tracking during nighttime encounters not only the common challenges of motion blur,object occlusion and scale variation but also inadequate lighting conditions and light interference.To improve the performance of real-time visual object tracking algorithms in low-light conditions and extend the application range of visual object tracking technology,this paper proposes a solution based on low-light image enhancement to address the problems of object tracking in nighttime.A low-light image enhancement module is used in the image preprocessing stage,combined with a real-time tracker for tracking.This paper proposes two low-light image enhancement algorithms based on supervised learning and unsupervised learning.The primary research accomplishments are delineated as follows.Firstly,we propose a low-light image enhancement method RAUNet based on Retinex theory.In response to poor lighting conditions in nighttime images,this paper proposes a reflection map estimation network RAUNet based on Retinex theory,which uses the reflection map as the result of low-light image enhancement.The method is based on the U-Net structure,and introduces an attention module to extract more effective features and reduce the impact of noise.Based on deep supervision,a multi-scale supervision module was designed to promote rapid convergence of network training and generate higher quality images.The test results of combining real-time tracker on public datasets show that sending images enhanced by RAUNet into the tracker can improve the success rate and precision of tracking.Secondly,we propose a low-light image enhancement method RDEnlighten GAN based on generative adversarial networks.RAUNet training network based on supervised learning requires image pairs composed of low-light and normal-light images in the same scene,and such image pairs are not easy to obtain.In addition to low-light condition in nighttime,the interference of lighting can lead to uneven lighting in the image,resulting in the loss of some features.In response to the above issues,this paper proposes a low-light image enhancement method RDEnlighten GAN based on generative adversarial networks.The generator combines the improved residual dense block to extract more effective local features,and uses the attention module of self-regularization to effectively keep the brightness of the image uniform.Employ both global and local discriminators to ensure the authenticity of details in both global and local regions.The experimental results of combining real-time tracker on nighttime visual object tracking datasets demonstrate the effectiveness of the RDEnlighten GAN method.In summary,in response to the problems of insufficient and uneven lighting in nighttime object tracking,this paper proposes a low-light image enhancement method based on Retinex theory and a low-light image enhancement network based on generative adversarial networks.The effectiveness of the two methods proposed in this paper was verified by testing them on public nighttime visual object tracking datasets using real-time tracker.更多还原