基于增强群跟踪器和深度学习的目标跟踪  

Target Tracking Based on Enhanced Flock of Tracker and Deep Learning

作　　者：程帅[1] 曹永刚[1,2] 孙俊喜[3] 赵立荣[1,2] 刘广文[1] 韩广良[2] 

机构地区：[1]长春理工大学电子信息工程学院,长春130022 [2]中国科学院长春光学精密机械与物理研究所,长春130000 [3]东北师范大学计算机科学与信息技术学院,长春130117

出　　处：《电子与信息学报》2015年第7期1646-1653,共8页Journal of Electronics & Information Technology

基　　金：国家自然科学基金(61172111);吉林省科技厅项目(20090512;20100312)资助课题

摘　　要：为解决基于外观模型和传统机器学习目标跟踪易出现目标漂移甚至跟踪失败的问题,该文提出以跟踪-学习-检测(TLD)算法为框架,基于增强群跟踪器(Fo T)和深度学习的目标跟踪算法。Fo T实现目标的预测与跟踪,增添基于时空上下文级联预测器提高预测局部跟踪器的成功率,快速随机采样一致性算法评估全局运动模型,提高目标跟踪的精确度。深度去噪自编码器和支持向量机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。加权P-N学习对样本加权处理,提高分类器的分类精确度。与其它跟踪算法相比较,在复杂环境下,不同图片序列实验结果表明,该算法在遮挡、相似背景等条件下具有更高的准确度和鲁棒性。To solve the problem that the tracking algorithm often leads to drift and failure based on the appearance model and traditional machine learning, a tracking algorithm is proposed based on the enhanced Flock of Tracker （FoT） and deep learning under the Tracking-Learning-Detection （TLD） framework. The target is predicted and tracked by the FoT, the cascaded predictor is added to improve the precision of the local tracker based on the spatio-temporal context, and the global motion model is evaluated by the speed-up random sample consensus algorithm to improve the accuracy. A deep detector is composed of the stacked denoising autoencoder and Support Vector Machine （SVM）, combines with a multi-scale scanning window with global search strategy to detect the possible targets. Each sample is weighted by the weighted P-N learning to improve the precision of the deep detector. Compared with the state-of-the-art trackers, according to the results of experiments on variant challenging image sequences in the complex environment, the proposed algorithm has more accuracy and better robust, especially for the occlusions, the background clutter and so on.

关 键 词：计算机视觉 群跟踪器 跟踪-学习-检测 深度学习 支持向量机 深度检测器 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]