【作者】 张宏达；

【导师】 孙统风；

【作者基本信息】 中国矿业大学 ， 计算机技术（专业学位）， 2020， 硕士

【摘要】 智能交通系统与日常生活联系密切,不仅关系到人身安全问题而且在个人和国家的发展中都占有重要地位。目标跟踪作为计算机视觉领域的基本问题之一,在智能交通系统中也有很大的实用价值。针对道路交通环境中背景复杂多变和快速运动等情况,本文在现有的相关滤波跟踪算法的基础上,从尺度估计和特征提取两个角度出发设计了适用于车辆的目标跟踪算法。第一,针对车辆跟踪场景中由快速运动引起的尺度变化剧烈的问题,提出了基于预测的尺度自适应相关滤波跟踪算法。该算法在现有相关滤波跟踪算法的多尺度跟踪策略前加入尺度预测环节,先使用前向差分运算得到预测尺度,然后在其附近构建尺度池并选取响应值最大的尺度作为目标尺度。随后利用反馈控制原理动态调整下一帧中的预测尺度,最终形成闭环控制提高跟踪算法尺度预测的精度。同时,该方法根据目标移动速度自适应更新跟踪模板,可以避免模型在线学习过程中学习到过量的背景信息。在KITTI和OTB100数据集上的实验结果表明该算法能够很好地适应快速运动场景,满足车辆目标跟踪的需求,与4种经典跟踪器相比精度和成功率得到了明显提升。第二,从特征角度出发设计了结合深层神经网络的相关滤波跟踪算法,以便挖掘更有效的特征从而提高跟踪器在复杂多变的交通环境下的判别力。该算法的特点在于使用了一种新的特征提取网络,通过剪除受padding影响的特征区域,缓解网络加深过程中padding对平移不变性的影响,从而使卷积神经网络中的深层特征可以有效应用于相关滤波跟踪算法中。在KITTI和OTB100数据集上的实验结果表明该模型可以有效发挥深层特征的能力,与5种经典跟踪器相比,该方法能够更好的应对车辆跟踪场景。更多还原

【Abstract】 Intelligent transportation systems are closely related to daily life,not only related to personal safety issues,but also occupy an important position in the development of individuals and countries.Target tracking,as one of the basic problems in the field of computer vision,has great practicality in intelligent transportation systems.Based on the existing tracking algorithms,two tracking algorithms suitable for vehicle targets are designed from the perspectives of scale and features to address the problems of complex and changeable backgrounds and fast motion in road traffic.First,a prediction-based adaptive correlation filter tracking algorithm based on scale is proposed to cope with the problem of rapid scale changes caused by rapid motion in vehicle tracking scenarios.The algorithm adds the scale prediction link to the multi-scale tracking strategy of the existing correlation filter tracking algorithm.First,the forward difference operation is used to obtain the prediction scale,and then the scale pool is built around it and the scale with the largest response value is selected as the target scale.Then use the feedback control principle to dynamically adjust the prediction scale in the next frame,and finally form a closed-loop control to improve the accuracy of the tracking algorithm scale prediction.At the same time,this method adaptively updates the tracking template according to the target moving speed,which can avoid excessive background information learning in the model online learning process.The experimental results on the KITTI and OTB100 data sets show that the algorithm can be well adapted to fast-moving scenes,and its accuracy and success rate are significantly improved compared with the four classic trackers.Second,from the perspective of features,a correlation filtering tracking algorithm combined with deep neural network is designed in order to mine more effective features and improve the discriminatory ability of the tracker in a complex and changing traffic environment.The feature of this algorithm is that it uses a new feature extraction network,which cuts off the feature areas affected by padding to mitigate the effect of padding on translation invariance during the network deepening process,so that the deep features in the convolutional neural network can be effectively applied In the correlation filter tracking algorithm.The experimental results on the KITTI and OTB100 data sets show that the model can effectively utilize the capabilities of deep features,and the method can better cope with vehicle tracking scenarios compared with 5 classic trackers.更多还原