复杂条件下运动目标跟踪方法的研究

发布时间：2018-04-18 14:36

本文选题：目标跟踪 + 跟踪-学习-检测　；参考：《长春理工大学》2016年博士论文

【摘要】：运动目标跟踪是计算机视觉的重要研究内容,在智能监控、精确制导、视觉导航等领域具有广泛的应用。尽管国内外研究人员对此进行了深入研究,但运动目标跟踪依然面临若干挑战性问题,突出表现在复杂条件下的目标姿态变化、光照变化、尺度变化、背景复杂、遮挡等因素的干扰。基于此,本文重点研究基于机器学习的目标跟踪算法,主要的研究工作如下:(1)提出了基于二值化规范梯度的TLD目标跟踪算法。在跟踪器中引入基于时空上下文的局部跟踪器失败预测方法和全局运动模型评估算法,提高跟踪精度。二值化规范梯度算法取代滑动窗口搜索策略,结合级联分类器实现目标检测,在保证检测精度的前提下,减少检测器的检测范围,提高了检测速度。将训练样本权重整合到在线学习过程中,提高级联分类器的分类精度,解决目标漂移问题。实验结果表明,与目前主流跟踪算法比较,在复杂条件下该算法提高跟踪精度的同时具有更快的处理速度。(2)提出了一种基于深度学习的目标跟踪算法。该算法以TLD算法为框架,利用增强群跟踪器对目标进行预测与跟踪,提高目标跟踪的精度。P-N学习对样本加权处理,提高分类器的分类精度。深度去噪自编码器和sigmoid分类器构建深度检测器,结合全局多尺度扫描窗口搜索策略检测可能的目标。深度去噪自编码器利用无监督特征学习优化网络参数,迁移学习将学习到的信息应用到在线跟踪中,提取图像的本质特征,同时解决跟踪过程中训练样本不足的问题。在线跟踪过程中,利用分类神经网络将目标从背景中分离出来,利用有监督学习微调网络参数,以适应跟踪过程中的各种变化。使用K均值聚类算法对在线模板集聚类,形成二值树,减少模板匹配数量,从而降低算法复杂度。实验结果表明,与目前主流跟踪算法比较,在复杂条件下该算法具有更高的跟踪精度及更好的稳定性。(3)提出了基于增量深度学习的目标跟踪算法。粒子滤波算法分布粒子集,通过无监督特征学习得到的深度去噪自编码器提取粒子区域特征,表达图像的本质信息。增量特征学习优化粒子区域特征集以适应目标的各种变化并实现粒子区域更有效的表达。增量特征学习由添加特征和整合特征两部分构成,添加特征引入新的特征信息,整合相似特征获得精简的特征表示。线性支持向量机对优化后的特征集进行分类,得到粒子置信度,同时微调深度网络,将粒子置信度最高的作为跟踪结果。引入粒子集规模自适应调整的双重采样过程,解决粒子衰减和贫化问题。实验结果表明,与目前主流跟踪算法比较,在复杂条件下该算法具有更高的跟踪精度以及更好的稳定性。(4)针对复杂条件下基于多示例学习的跟踪算法存在的问题,提出了基于多示例深度学习的目标跟踪算法。针对原始多示例跟踪算法中采用Haar-like特征不能有效表达图像信息以及受到外界条件的影响很容易失效的缺点,利用深度去噪自编码器提取示例图像的有效特征,实现图像信息的本质表达,提高对环境变化的适应能力。针对原始多示例学习跟踪算法中选取弱特征向量不能更换,难以反映目标自身和外界条件变化的缺点,在选择弱分类器过程中,实时替换判别力最弱的特征以适应目标外观的变化。针对原始多示例跟踪算法中运动模型仅假设帧间物体运动不会超过某个范围,不能有效反映目标运动状态的缺点,引入粒子滤波算法对目标进行预测,提高跟踪精度。实验结果表明,与多示例学习跟踪算法及其他主流跟踪算法比较,在复杂环境下该算法提高了跟踪精度和跟踪的稳定性。
[Abstract]:Moving target tracking is an important research topic in computer vision, intelligent monitoring, precision guidance, has been widely applied to the field of visual navigation. Although the domestic and foreign researchers have conducted in-depth research, but the moving target tracking is still facing some challenging problems, highlighting the target attitude changes under complex conditions, illumination changes, scale change, complex background and occlusion interference factors. Based on this, this paper focuses on the research of machine learning based target tracking algorithm, the main research work is as follows: (1) proposed two value tracking algorithm standardized TLD target. Based on gradient prediction method and global motion model evaluation algorithm of local tracker spatiotemporal context failure based on the introduction of the tracker, improve the tracking accuracy. The binarization standard gradient algorithm instead of sliding window search strategy, the target detection based on cascade classifier Guarantee the detection accuracy and reduce the detection range of the detector, the detection speed is improved. The training sample weights integrated into the online learning process, to improve the classification accuracy of cascade classifier, solve the drift problem. The experimental results show that compared with the current mainstream tracking algorithms, and improve tracking accuracy with faster processing speed in algorithm the complex conditions. (2) proposed a tracking algorithm based on deep learning target. The algorithm uses the TLD algorithm as the framework, the use of enhanced group tracker prediction and tracking of targets, improve the accuracy of target tracking.P-N learning to sample weighting processing, improve classification accuracy. The depth of self denoising encoder and sigmoid classifier construction depth detector, combining the global multi-scale scanning window search target detection strategy possible. Since the depth of denoising using unsupervised feature encoder Learning to optimize the network parameters, transfer learning will learn information applied to online tracking, the essence of image feature extraction, and solve the problems in the process of tracking samples. In the process of online tracking, classification using neural network to separate the object from the background, the use of supervised learning to fine tune the network parameters to adapt to the change in the tracking process. Clustering on online template using K clustering algorithm, the formation of two value tree, reduce the number of template matching, so as to reduce the complexity of the algorithm. The experimental results show that compared with the current mainstream tracking algorithm, tracking accuracy and stability is better in the algorithm under complex conditions (3) is proposed. The tracking algorithm of incremental deep learning targets based on particle filter algorithm. The particle distribution, through unsupervised learning characteristics of the depth of denoising from encoder regional feature extraction particle, Express the essential information of the images. The optimization of regional characteristics of the particle set to adapt to the change of target and realize regional expression of particle more effective incremental learning characteristics. Incremental feature learning by adding features and characteristics of the integration of two parts, adding new features into feature information integration, said similar features obtained streamlined linear support set on the characteristics of features. The optimized vector machine classification, particle confidence, and fine tune the depth of network, the particle will be the highest confidence as the tracking results. Double sampling process is introduced into the particle set scale adaptive, solve the particle decay and dilution. The experimental results show that compared with the current mainstream tracking algorithms, has higher tracking accuracy and better stability in the algorithm under complex conditions. (4) according to the existing tracking algorithm of multi instance learning based on problems under complex conditions, put forward The tracking algorithm based on multi instance deep learning goal. According to the original multi sample tracking using Haar-like features can not express image information and affected by external conditions easily failure algorithm, using depth denoising from encoder sample image effective feature extraction, realize the essence of image information expression, improve the change of environment the ability to adapt. For the replacement of weak feature vectors can not select the original miltracking algorithm, it is difficult to reflect the change of the target itself and external conditions in the choice of the shortcomings of weak classifier in the process of replacing the weakest feature of real-time discrimination in order to adapt to the changes of appearance of objects. According to the original sample tracking algorithm in motion model assumes only inter object exercise does not exceed a certain range, can not effectively reflect the moving state of the target faults into the particle filter algorithm in pre target Experimental results show that compared with multiple instance learning tracking algorithm and other mainstream tracking algorithms, the algorithm improves tracking accuracy and tracking stability in complex environment.

【学位授予单位】：长春理工大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】