【作者】 吴磊；

【导师】 夏威；

【作者基本信息】 电子科技大学 ， 电子信息（专业学位）， 2023， 硕士

【摘要】 近年来,伴随着计算机数据处理能力的增强,多目标跟踪在智能安防、人机交互、视频分析等多个领域具有广阔的应用前景。行人多目标跟踪的主要任务是在视频序列中对目标行人进行持续性的检测与定位,并完成帧间目标的关联。该任务易受到视角、光照、遮挡等干扰因素的影响。在目前以深度学习为基础的多目标跟踪方法中,联合检测与跟踪的方法将行人检测任务和行人跟踪任务集成到同一个端到端神经网络中,其优势是具有更快的跟踪速度,缺点则是当遇到遮挡、密集场景等情况时,跟踪精度会出现明显的下降。针对以上问题,本文的具体工作内容如下:（1）针对多目标跟踪任务中行人检测边界框估计不准以及遮挡等干扰问题的影响,本文设计了一种基于目标检测与运动估计的行人多目标跟踪方法。首先设计了以34层残差网络为骨干,集成多层特征聚合模块和多任务分支的行人检测模型,对视频帧中的每一个行人进行检测定位;然后采用修改后的卡尔曼滤波器建立行人的运动估计模型,对下一视频帧中的行人位置进行预测;最后以匈牙利算法为基础,建立基于多阶段数据关联的检测—轨迹匹配算法,完成前后视频帧中同一身份行人的关联跟踪。该行人多目标跟踪方法在MOT16和MOT17测试集上的MOTA分别达到了75.6%和74.3%,IDF₁达到了73.1%和72.4%。（2）在（1）中设计的行人多目标跟踪方法的基础上,本文设计了一种联合重识别特征辅助和多阶段数据关联的行人多目标跟踪方法,以提高行人身份的关联能力。该方法引入单独训练的行人重识别模型以提取行人的重识别特征,并将该特征添加到多阶段数据关联算法中提高多目标跟踪的准确率。重识别模型以ResNeSt50为骨干网络,通过单独的卷积结构进行特征降维,并在MOT17数据集上完成独立训练。添加重识别特征辅助后的多目标跟踪方法在MOT16和MOT17两个测试集上均展现出了更强的身份关联能力,IDF₁较（1）中方法分别提升了1.9%和2.1%。综上所述,本文设计了一种先检测再跟踪的多目标跟踪算法,通过在目标检测模型、运动估计模型以及数据关联方法上的改进,在具有较高跟踪精度的同时保障了跟踪速度。重识别模型的添加进一步提高了对行人身份的关联能力。本文通过一系列消融实验证明了各个模块的有效性,同时在MOT Challenge上进行的在线测试表明,本文设计的方法优于一些主流的行人多目标跟踪方法。更多还原

【Abstract】 In recent years,multi-target tracking has gained increasing attention in various fields,including intelligent security,human-computer interaction,and video analysis due to the advancement of computer data processing capabilities.The primary goal of this task is to detect and locate targets in video sequences and associate targets between frames.Deep learning-based approaches for multi-target tracking have been developed,such as the joint detection and tracking approach.This method combines pedestrian detection and tracking tasks into a single neural network,the advantage is faster tracking speed,but the disadvantage is that when encountering situations such as blocked targets or excessively dense targets,the tracking accuracy can be significantly reduced.（1）To address the effects of inaccurate estimation of pedestrian detection bounding boxes and interference problems such as occlusion in multi-target tracking tasks,thesis proposes a method for tracking multiple pedestrians using a combination of target detection and motion estimation.Firstly,a pedestrian detection model with 34 layer residual network as the backbone,integrated multi-layer feature aggregation module and multi-task branching is designed to detect and locate each pedestrian in a video frame.Then a modified Kalman filter is used to build a pedestrian motion estimation model to predict the pedestrian position in the next video frame.Finally,based on the Hungarian algorithm,a detection-trajectory matching algorithm based on multi-stage data association is established to complete the association tracking of the same identity pedestrian in the before and after video frames.The MOTA of this pedestrian multi-target tracking method on the MOT16 and MOT17 test sets reached 75.6%and 74.3%,respectively,while the IDF₁ reached 73.1%and 72.4%.（2）Based on the pedestrian multi-target tracking method designed in（1）,thesis designs a pedestrian multi-target tracking method that combines recognition feature assistance and multi-stage data association to improve the ability to correlate pedestrian identities.The method introduces a separately trained pedestrian reidentification model to extract the reidentification features of the pedestrian and adds the features to the multi-stage data association algorithm to improve the accuracy of multi-target tracking.The re-recognition model uses ResNeSt50 as the backbone network,performs dimensionality reduction of the features through separate convolution structures and is trained on the MOT17 multi-target tracking dataset,independently.The multi-target tracking method assisted by the addition of re recognition features exhibits stronger pedestrian identity association capabilities on both MOT16 and MOT17 test sets,with IDF₁increasing by 1.9%and 2.1%compared to the method in（1）,respectively.In summary,thesis designs a detection-by-tracking multi-target tracking algorithm,which achieves excellent tracking accuracy while ensuring tracking speed by improving the target detection model,motion estimation model and data association method.The addition of a recognition model further improves the pedestrian identity association ability.Thesis has demonstrated the effectiveness of each module through a series of ablation experiments,and online testing on MOT Challenge has shown that the method designed in thesis is superior to some mainstream pedestrian multi-target tracking methods.更多还原