结合相关滤波和卷积神经网络的目标跟踪方法研究

发布时间：2018-05-29 23:25

本文选题：特征提取 + 卷积神经网络　；参考：《西北农林科技大学》2017年硕士论文

【摘要】：目标跟踪算法是计算机视觉中用于处理视频图像信息的重要技术手段。在目标跟踪的过程中要面对目标快速移动、背景复杂、遮挡、光照变化等不定因素的挑战,对跟踪算法鲁棒性要求高。近来,将深度学习与目标跟踪算法结合成为目标跟踪研究领域的热点。本文把卷积神经网络与相关滤波算法结合作为研究对象。通过对卷积神经网络分层研究,深入的分析各层结构的特征性质。为了满足跟踪算法的实时性,本文对原VGG-Net网络结构进行调整,使用改进后的卷积神经网络对图像目标进行特征提取。同时,针对跟踪算法的尺度自适应及目标模型更新进行了改进。本文的主要内容如下所示:(1)领域自适应卷积神经网络特征提取。通过对卷积神经网络提取的特征进行分析研究,将深度特征与核相关滤波跟踪算法进行结合。本研究采用结构调整后的VGG-Net对目标样本进行特征提取,卷积神经网络中靠前的卷积层提取到的特征保留较多的空间信息,靠后的卷积层提取的特征具有目标语义信息。算法中使用三层的卷积特征训练滤波器,采用由粗到细的方法,将三个滤波器跟踪结果进行综合,实现对跟踪目标的精确定位。(2)模型自适应更新策略改进。大部分目标跟踪算法中,每一帧目标跟踪定位结束后,通过模型更新机制对滤波器模型进行更新,更新的过程需要再次对目标样本进行特征提取,导致跟踪算法变慢。同时,因为跟踪目标存在被同类物体或背景遮挡的现象,在跟踪结果准确度低的情况下,对滤波器模型更新会导致模型污染,造成跟踪漂移,最终导致跟踪失败。为了减少计算负担,提高跟踪精度,本算法在模型更新模块中使用了最大响应值和平均相对峰值两种更新指标,当跟踪结果符合更新条件时才对模型进行更新。(3)目标尺度自适应。目标跟踪过程中,跟踪目标与摄像头的相对移动会导致目标在图像中尺度的变化,如果算法不能自适应尺度变化,在检测过程中会导致跟踪结果漂移,本文通过增加单独的尺度滤波器对跟踪目标的尺度变化进行实时的估计,进而实时调整样本采样框大小,避免跟踪精度受尺度变化影响。
[Abstract]:Target tracking algorithm is an important technique for processing video image information in computer vision. In the process of target tracking, we have to face the challenges of fast moving target, complex background, occlusion, illumination change and other uncertain factors, so the robustness of tracking algorithm is very high. Recently, the combination of depth learning and target tracking algorithm has become a hotspot in the field of target tracking. In this paper, the convolutional neural network and the correlation filtering algorithm are combined as the research object. Through the hierarchical research of convolutional neural network, the characteristic properties of each layer structure are analyzed in depth. In order to meet the real-time performance of the tracking algorithm, the original VGG-Net network structure is adjusted and the improved convolution neural network is used to extract the feature of the image target. At the same time, the scale adaptation and target model updating of tracking algorithm are improved. The main contents of this paper are as follows: 1) Domain adaptive convolution neural network feature extraction. By analyzing the feature extracted by convolution neural network, the depth feature is combined with the kernel correlation filter tracking algorithm. In this study, VGG-Net is used to extract the features of target samples. The features extracted from the convolution layer in the convolution neural network retain more spatial information, and the features extracted from the convolution layer have the semantic information of the target. In the algorithm, the three-layer convolution feature training filter is used, the coarse-to-fine method is used to synthesize the tracking results of the three filters, and the adaptive updating strategy for the accurate location of the target is realized. In most of the target tracking algorithms, the filter model is updated by model updating mechanism after each frame target tracking is finished, and the updating process needs to extract the feature of the target sample again, which results in the slow down of the tracking algorithm. At the same time, because of the phenomenon that the target is blocked by the same object or background, the model updating of filter will lead to model pollution, tracking drift, and eventually lead to tracking failure when the accuracy of the tracking results is low. In order to reduce the computational burden and improve the tracking accuracy, this algorithm uses two updating indexes, the maximum response value and the average relative peak value, in the model updating module. When the tracking results meet the updating condition, the model is updated. In the process of target tracking, the relative movement of the target and the camera will lead to the change of the target scale in the image. If the algorithm can not adapt the scale change, it will lead to the drift of the tracking result in the detection process. In this paper, a single scale filter is added to estimate the scale change of the tracking target in real time, and then the size of the sample sampling frame is adjusted in real time to avoid the tracking accuracy affected by the scale change.
【学位授予单位】：西北农林科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】