基于级联特征的随机森林运动目标跟踪算法-《计算机技术与发展》

文章信息/Info

Title:: A Moving Object Tracking Algorithm of Random Forest Based on Features Cascade

作者:: 陆兵¹; 顾苏杭¹; 2; 1. 常州轻工职业技术学院信息工程学院,江苏常州 213164;2. 江南大学数字媒体学院,江苏无锡 214122

Author(s):: LU Bing¹; GU Su-hang¹; 2; 1. School of Information Engineering,Changzhou Vocational Institute of Light Industry,Changzhou 213164,China;2. School of Digital Media,Jiangnan University,Wuxi 214122,China

Keywords:: complex environment; cascading features; contour; random forest; positive and negative sample

摘要:: 在运动目标检测与跟踪的过程中,实际环境下的目标旋转、目标遮挡以及光照变化等因素时常出现,而目标检测与跟踪的性能对这些复杂环境因素极为敏感,甚至易导致目标跟踪丢失。为了提高复杂环境下运动目标跟踪的鲁棒性和稳定性,提出一种基于级联特征的随机森林运动目标跟踪算法。该算法首先在保留目标关键信息的 ASIFT 特征中级联目标轮廓信息作为正样本,训练正样本生成随机森林分类后续序列图像特征;在此基础上将 CamShift 算法确定的目标搜索窗口中的非目标特征作为负样本,训练负样本并更新随机森林以改善特征分类性能;最后通过对正负样本特征加权计算目标搜索窗口质心以改善跟踪性能。实验结果表明,该算法能够在光照突变、遮挡以及目标旋转等复杂环境下有效地实现运动目标跟踪。

Abstract:: In the process of moving object detection and tracking,in actual scenarios where there usually exists complex environmental factors including object rotation,occlusion and illumination and so on, the performance of the object detection and tracking is easily affected by these complicated environmental factors,even they lead to the occurrence of the losses of object tracking. In order to improve the robustness and stability of moving object tracking under complex environment,we propose a random forest for moving object trackingalgorithm based on features cascade. The ASIFT features of the moving object with retaining key information are cascaded with objectcontour information as positive sample set. Random forest which can be used to classify the features of the subsequent sequence images isrealized through training the positive sample set. On the basis of cascading features,the features of non - object in the object searchwindow determined by CamShift algorithm are taken as negative sample set,and the performance of feature classification is improved bytraining negative sample set which is used to update random forest. The centroid of the object search window is calculated by weightbased positive and negative sample sets to improve the tracking performance. The experiment indicates that the algorithm can effectivelyrealize moving object tracking under complex environment such as illumination fluctuation,occlusion and object rotation.