侯建华,边群星,项 俊.基于在线学习判别性外观模型的多目标跟踪算法[J].中南民族大学学报自然科学版,2017,(1):81-86
基于在线学习判别性外观模型的多目标跟踪算法
Multi-Object Tracking Algorithm Based on Online Learned Discriminative Appearance Model
  
DOI:
中文关键词: 多目标跟踪  轨迹片  外观模型  Adaboost 算法
英文关键词: multi-object tracking  tracklets  appearance model  Adaboost algorithm
基金项目:国家自然科学基金资助项目(61671484);中南民族大学中央高校基本科研业务费专项资金(CZW15013)
作者单位
侯建华,边群星,项 俊 中南民族大学 电子信息工程学院武汉 430074 
摘要点击次数: 261
全文下载次数: 260
中文摘要:
      在基于检测的跟踪框架下,设计了一种在线学习的判别性外观模型并应用于多目标跟踪.对检测器输出 的相邻帧间检测响应做保守关联,生成短小可靠的轨迹片; 利用目标轨迹时空域约束条件,从轨迹片中提取训练样 本及特征; 采用 Adaboost 算法在线生成目标外观的判别性模型,计算轨迹片之间的外观相似度; 最后采用匈牙利算 法,经过多次迭代得到每个目标的完整轨迹. 对实验结果做了定量和定性分析, 结果表明: 所设计的算法提高了跟 踪精度,在复杂场景下能够较好地完成多目标跟踪任务.
英文摘要:
      An online learned discriminative appearance models is designed for multi-object tracking ( MOT) under the tracking by detection framework. Firstly, by low level association strategy, short yet reliable tracklets are generated among detection responses which are produced by a pre-trained detector. Then, the training samples are collected by using spatialtemporal constraints of target trajectory, and features are extracted for representation of the appearance model. We adopt online Adaboost algorithm to train the discriminative appearance model, by which the appearance similarity between tracklets can be calculated. Finally, the Hungarian algorithm is used via several iterations to obtain the complete trajectory for each target. Experiments are conducted quantitatively and qualitatively. The results show that the proposed method has improved the tracking accuracy, and has satisfactory performance in complex scene.
查看全文   查看/发表评论  下载PDF阅读器
关闭