【作者】 王冠；

【导师】 宫俊；

【作者基本信息】 东北大学 ， 控制工程（专业学位）， 2021， 硕士

【摘要】 目标跟踪是计算机视觉领域重要的研究方向之一,是集特征提取、目标检测和动态分析于一体的综合性课题,在无人驾驶、智能监控以及军事侦察等领域得到了广泛应用,具备极高的科研意义和实用价值。近年来目标跟踪领域的研究方法层出不穷,但仍然面临着诸多影响因素的挑战,如形变、快速运动、尺度变化和遮挡等,跟踪模型的准确度及鲁棒性仍有待提高。目前基于孪生网络的目标跟踪模型发展迅速,逐渐成为该领域的主流研究方向。本文基于SiamRPN(孪生区域推荐网络)模型,针对目标跟踪领域存在的目标尺度变化和形变问题,提出一系列优化策略和改进方案,主要研究内容如下:(1)针对SiamRPN模型的特征提取能力不足和正负样本不均衡的问题,本文在基准网络与损失函数两方面作出优化,构造以VGG-Net-16为主干的特征提取基准网络来替换原有的AlexNet,在分类损失函数中引入收缩项来抑制简单负样本,添加以IoU为纽带的耦合因子来联合分类任务和边框回归任务,增加分类任务和边框回归任务的耦合性,提高了模型的分类判别能力和边框回归精度;(2)针对跟踪过程中存在的目标尺度变化问题,本文构造了一种融合注意力机制与级联RPN的SiamRPN模型。基于注意力机制的原理构造出双重互联注意力模块,采用级联RPN的形式处理多个卷积层提取到的不同尺度特征信息,充分利用了目标的高层语义信息以及底层纹理信息协同推断目标分类与定位,提高了目标定位的准确度,增加了模型对于目标尺度变化的鲁棒性;(3)针对跟踪过程中存在的目标形变问题,本文构造了一种引入模板自适应更新策略的SiamRPN模型。提出了一种模板自适应更新策略,在目标跟踪过程中基于特征图响应动态决策是否更新原目标模板,当满足更新判定条件时利用模板更新子网络生成最新的目标模板响应图,解决因目标形状变化导致的模板信息匮乏问题,从而提升了模型的跟踪精度以及对于形变问题的适应能力;(4)本文提出的模型经过离线训练后,在OTB100和VOT2018数据集上作不同维度上的定量测试评估,实验结果充分验证了本文优化策略和改进方案的有效性,本文模型的性能指标相较于其他主流跟踪模型处于领先地位。从OTB100数据集中挑选具有不同挑战因素的视频序列与当下流行的跟踪模型作定性测试,可以直观地展示本文模型与其他模型跟踪效果的差异性。此外,本文模型的运行速度均能满足实时性的要求。更多还原

【Abstract】 Object tracking is one of the important research directions in the field of computer vision.It is a comprehensive subject that integrates feature extraction,object detection and dynamic analysis.It has been widely used in the fields of unmanned driving,intelligent monitoring and military reconnaissance,and has extremely high scientific research significance and practical value.In recent years,research methods in the field of object tracking have emerged one after another,but they are still facing challenges from many influencing factors,such as deformation,rapid motion,scale variation and occlusion.The accuracy and robustness of the tracking model still needs to be improved.At present,the object tracking model based on the siamese network has developed rapidly and has gradually become the mainstream research direction in this field.Based on the SiamRPN model,this paper proposes a series of optimization strategies and improvement schemes for scale variation and deformation problems in the object tracking.The main research contents are as follows:(1)In view of the Insufficient feature extraction capability and the imbalance of positive and negative samples in the SiamRPN model,this paper optimizes the backbone network and loss function,and constructs a feature extraction benchmark network based on VGG-Net-16 to replace the original AlexNet.In the classification loss function,the contraction item is introduced to suppress simple negative samples,and the coupling factor with the IoU is added to combine classification and bounding box regression task to increase the coupling between classification task and bounding box regression task.It improves the discriminative ability and regression accuracy of the model.(2)Aiming at the problem of object scale variation in the tracking process,this paper designs a SiamRPN model that combines the attention mechanism and the cascaded RPN.Based on the principle of the attention mechanism,a dual interconnected attention module is constructed,and the cascaded RPN is used to process the feature information of different scales extracted by multiple convolution layers.It makes full use of the high-level semantic information of the object and the low-level texture information to collaboratively infer the target classification and position,improves the accuracy of object positioning and increases the model’s robustness to object scale variation.(3)Aiming at the problem of object deformation in the tracking process,this paper designs a SiamRPN model that introduces template adaptive update strategy.The template adaptive update strategy is proposed.In the object tracking process,the model dynamically decide whether to update the original target template based on the feature map response.When the update judgment condition is met,the template update sub-network is used to generate the latest object template response,which solves the problem of lack of template information caused by the change of object shape,thereby improves the tracking accuracy of the model and the ability to adapt to the problem of deformation.(4)After offline training,the tracking model proposed in this paper is quantitatively tested and evaluated on the OTB100 and VOT2018 datasets in different dimensions.The results of comparative experiments fully verify the effectiveness of the optimization strategy and improvement scheme in this paper.Compared with the other mainstream tracking model,the performance index of the model in this paper is in a leading position.Select video sequences with different challenge factors from the OTB100 dataset and do qualitative tests with the current popular object tracking models,the test results can intuitively show the difference between the tracking effect of this model and other models.In addition,the running speed of the model in this paper can meet the real-time requirements.更多还原