基于在线学习判别性外观模型的多目标跟踪算法.pdf

第 36卷第1期 中南民族大学学报（自然科学版） Yol . 36 No . 1 2017年 3 月 Journal of South-Central University for Nationalities( Nat. Sci. Edition) Mar . 2017 基于在线学习判别性外观模型的多目标跟踪算法 侯建华，边群星，项 俊 ( 中南民族大学电子信息工程学院，武汉43 0074 ) 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A dab oost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. 关键词多目标跟踪 ; 轨迹片; 外观模型; A dab oost算法 中 图 分 类 号T P 3 9 1 .41 文献标识码 A 文 章 编 号 1 6 7 2 4 3 2 1 (2 0 1 7 )0 1 -0 0 8 1 -0 6 Multi-Object Tracking Algorithm Based on Online Learned Discriminative Appearance Model Hou Jianhua , Bian Qunxing , Xiang Jun ( College o f Electronic Inform ation Engine erin g, South-C entral U niversity fo r N a tio n a litie s, W uhan 43 0 0 7 4 ) A b s tra c t A n online learned discrim inative appearance models is designed fo r m u lti-o b je ct tra cking ( M O T ) under the tra cking by detection fram ew ork. F irs tly , by low level association strategy, short yet reliable tracklets are generated among detection responses w hich are produced by a pre-trained detector. T h e n , the tra in in g samples are collected by using spatial- tem poral constraints o f target tra je cto ry, and features are extracted fo r representation o f the appearance m odel. W e adopt online Adaboost algorithm to tra in the discrim inative appearance m o d e l, by w hich the appearance s im ila rity between tracklets can be calculated. F