【作者】 李钢；

【导师】 钟胜；

【作者基本信息】 华中科技大学 ， 控制工程， 2021， 硕士

【摘要】 目标跟踪技术是计算机视觉领域的一个重要分支,在安防监控、视觉导航、国防科技等领域有着广泛的应用,具有重要的研究意义。研究人员对目标跟踪技术进行了深入的研究,其中基于相关滤波的目标跟踪方法以其高效性在众多目标跟踪方法中脱颖而出。由于跟踪场景的复杂性和不确定性,基于相关滤波的目标跟踪方法在许多场景中如在无人机视角下目标快速运动、遮挡、出视野、视点运动时容易跟踪失败。本文针对相关滤波的特征融合、尺度估计、模型更新、抗遮挡等方面展开研究。针对相关滤波目标跟踪算法中使用固定融合系数对不同特征的相关滤波响应图进行融合,不能及时适应图像场景变化,以及尺度估计时计算量太大的问题,本文提出了基于自适应融合特征的分类搜索尺度估计的方法。动态分配不同特征的融合权重,发挥不同特征的优势;尺度估计时判断尺度变化趋势,减小尺度搜索范围,从而减少计算量。在UAV123数据集上的实验结果表明,所提方法的精确度和成功率相比原始算法分别提高了3.7%和8.9%,平均帧频提高了12.6%。针对相关滤波目标跟踪算法使用固定学习率更新模型易导致模型污染,本文提出了自适应学习率跟踪方法,判断目标跟踪效果,动态调整学习率。同时本文提出了基于目标与背景联合建模的目标跟踪方法。在目标快速运动或遮挡时使用模板匹配和最小二乘法估计目标位置,从而解决该场景容易跟踪漂移或失败的问题。结合这两种方法,在UAV123数据集上的实验结果表明,跟踪的精确度和成功率相比原始算法分别提高了27.4%和15.7%。本文在华为Atlas 200 DK嵌入式开发套件上实现了本文提出的算法。针对无人机平台跟踪算法的实时性和低功耗要求,结合ARM架构采用了指令集优化和多线程并行,提高算法执行的效率。实验结果表明,本文算法在开发板上对于每个测试数据集序列的平均帧频都大于25帧,满足实时性的要求。开发板运行跟踪算法时的整板功耗为7.116W,满足低功耗的要求,具有在无人机平台上的应用价值。更多还原

【Abstract】 Target tracking is an important topic in computer vision,which has made significant progress in the past few years.It has been widely applied in security monitoring,visual navigation,and national defense technology.Researchers have conducted in-depth research on target tracking,among which the correlation filtering-based target tracking method outperforms many target tracking methods for its high efficiency.However,target tracking based on correlation filtering has some shortcomings in many scenes because of the complexity and uncertainty of the tracking scene.For example,it is easy to fail to track when the target is moving fast,occluded,out of view,or in the UAV view with the viewpoint motion.This thesis mainly focuses on correlation filtering’s feature fusion,scale estimation,model updating,and anti-occlusion.The correlation filtering target tracking algorithm uses fixed fusion coefficients to fuse the correlation filter response maps of different features.This method cannot adapt to image scene changes in time during target tracking,and is too computationally complicated for scale estimation.To mitigate these problems,this thesis proposes an adaptive fusion features and classification search scale estimation method.The method dynamically assigns fusion weights to different features to take full advantage of different features,and also narrows the scale search range by judging the trend of scale change to reduce the computation complexity in the scale estimation.The experimental results on the UAV123 dataset show that the accuracy and success rate of the proposed method are increased by 3.7% and 8.9%compared with the original algorithm,and the average frame rate is increased by 12.6%.The correlation filtering target tracking algorithm uses fixed learning rate to update the model.This approach results in model pollution.To address the problem,this thesis proposes an adaptive learning rate tracking method.It first judges the effect of target tracking,and then dynamically tunes the learning rate.At the same time,this thesis proposes a joint modelling of target and background tracking method to solve the problem that target tracking is easy to track drift or failure when the target is moving fast or occluded.It uses the template matching method and the least square method to estimate the target position.Combining these two methods,the experimental results on the UAV123 dataset show that the accuracy and success rate of the proposed method are improved by 27.4% and 15.7%compared with the original algorithm.The algorithm proposed in this thesis is achieved on Huawei Atlas 200 DK.For the real-time and low power consumption requirements of UAV platform tracking algorithms,this thesis uses instruction set optimization and multi-thread parallelism with ARM architecture to improve the efficiency of algorithm execution.The experimental results show that the average frame rate of the algorithm on the development board for each test data set sequence is greater than 25 frames,which meets the requirement of real-time performance.The power consumption of the whole board when the development kit is running the tracking algorithm is 7.116 W,which meets the requirement of low power consumption and has the value of application in the UAV platform.更多还原