【作者】 高倩；

【导师】 宋焕生；

【作者基本信息】 长安大学 ， 计算机技术（专业学位）， 2023， 硕士

【摘要】 交通场景中监控相机的应用已经变得广泛普及,产生了大量的交通视频数据。这些视频数据需要进行深度信息挖掘,以实现更为深层次的应用。目前大多数交通视频分析研究都是针对单个监控相机的独立场景,没有充分利用相机间的时空关联进行综合分析和决策判断。本文针对实际应用场景中多目标跟踪算法效率低以及遮挡严重情况下目标跟踪精度差的问题,着重开展单相机下多目标跟踪与跨相机车辆跟踪技术研究。主要内容如下:(1)单相机车辆多目标跟踪技术研究。针对单相机下的多目标跟踪算法效率低,难以满足工程上实时性问题,设计了两种多目标跟踪算法。一种是基于光流和卡尔曼滤波的多目标跟踪算法,该算法通过设计了一种多维度约束的目标轨迹关联策略,利用中值光流算法快速获取车辆速度,结合卡尔曼滤波得到的目标预测位置,实现跨帧的多目标快速跟踪。另一种是基于Fair MOT的多目标跟踪算法,该算法通过分析Re ID分支获取的特征不稳定的原因,充分考虑通道维度上的不足,引入SE注意力机制,实现较高精度的多目标跟踪。通过实验验证了两种多目标跟踪算法的有效性,并进行对比分析。(2)隧道场景下跨相机跟踪数据集的构建。通过分析交通场景中公开的跨相机跟踪数据集的不足,设计了一种基于多目标跟踪的轨迹级标注方法。该方法利用单相机多目标跟踪算法得到轨迹初始标注结果,使用Ultimate Labeling标注工具对多个相机下的同一车辆进行关联标注。最终,构建了隧道场景下跨相机跟踪数据集,并与其他跨相机跟踪数据集进行对比分析。(3)隧道场景下跨相机车辆跟踪技术研究。针对隧道场景下跨相机车辆受光照,遮挡等导致跟踪精度较低的问题,提出了基于单相机跟踪轨迹的跨相机跟踪方案。该方案通过卡尔曼滤波进行轨迹预测获取车辆空间位置特征,结合基于SGE注意力机制的残差网络获取车辆重识别特征。然后将这两种特征作为轨迹关联线索,设计了一种多特征轨迹关联机制,实现隧道场景下车辆的跨相机大范围持续跟踪。本文研究的交通场景下跨相机跟踪技术可应用于交通监控系统中的车辆行为分析与监测,可以较好地解决大范围连续的车辆运动状态分析等问题。更多还原

【Abstract】 The application of surveillance cameras in traffic scenes has become widely popular,generating a large amount of traffic video data.These video data require in-depth information mining to achieve deeper applications.At present,most traffic video analysis researches focus on the independent scene of a single surveillance camera,and do not make full use of the temporal and spatial correlation between cameras for comprehensive analysis and decisionmaking.Aiming at the low efficiency of multi-target tracking algorithms in practical application scenarios and poor target tracking accuracy under severe occlusion,this paper focuses on the research of multi-target tracking under single camera and cross-camera vehicle tracking technology.The main contents are as follows:(1)Research on single-camera vehicle multi-target tracking technology.Aiming at the low efficiency of multi-target tracking algorithm under single camera,it is difficult to meet the realtime problem in engineering,and two multi-target tracking algorithms are designed.One is a multi-target tracking algorithm based on optical flow and Kalman filtering.This algorithm designs a multi-dimensional constraint target trajectory association strategy,uses the median optical flow algorithm to quickly obtain vehicle speed,and combines the target obtained by Kalman filtering.Predict the position and realize the fast tracking of multiple targets across frames.The other is the multi-target tracking algorithm based on Fair MOT.This algorithm analyzes the reasons for the instability of the features obtained by the Re ID branch,fully considers the lack of channel dimensions,and introduces the SE attention mechanism to achieve higher-precision multi-target tracking.The effectiveness of the two multi-target tracking algorithms is verified by experiments,and a comparative analysis is carried out.(2)Construction of cross-camera tracking dataset in tunnel scene.By analyzing the deficiencies of publicly available cross-camera tracking datasets in traffic scenes,a trajectorylevel annotation method based on multi-object tracking is designed.This method uses a singlecamera multi-target tracking algorithm to obtain the initial labeling result of the trajectory,and uses the Ultimate Labeling labeling tool to label the same vehicle under multiple cameras.Finally,a cross-camera tracking dataset in tunnel scenes was constructed and compared with other cross-camera tracking datasets.(3)Research on cross-camera vehicle tracking technology in tunnel scene.Aiming at the problem of low tracking accuracy caused by illumination and occlusion of cross-camera vehicles in tunnel scenes,a cross-camera tracking scheme based on single-camera tracking trajectory is proposed.The program uses Kalman filter to predict the trajectory to obtain the vehicle spatial position features,and combines the residual network based on the SGE attention mechanism to obtain the vehicle re-identification features.Then,these two features are used as trajectory association clues,and a multi-feature trajectory association mechanism is designed to realize the continuous tracking of vehicles across cameras in a tunnel scene.The cross-camera tracking technology studied in this paper can be applied to the analysis and monitoring of vehicle behavior in the traffic monitoring system,and can better solve the problems of large-scale continuous vehicle motion state analysis.更多还原