目的 针对基于检测的目标跟踪问题,提出一种联合多特征融合和判别性外观模型的多目标跟踪算法。方法 对时间滑动窗内的检测器输出响应,采用双阈值法对相邻帧目标进行初级关联,形成可靠的跟踪片,从中提取训练样本;融合多个特征对样本进行鲁棒表达,利用Adaboost算法在线训练分类器,形成目标的判别性外观模型;再利用该模型对可靠的跟踪片进行多次迭代关联,形成目标完整的轨迹。结果 4个视频数据库的目标跟踪结果表明,本文算法能较好的处理目标间遮挡、目标自身形变,以及背景干扰。对TUD-Crossing数据库的跟踪结果进行了定量分析,本文算法的FAF(跟踪视频序列时,平均每帧被错误跟踪的目标数)为0.21、MT(在整个序列中,有超过80%视频帧被跟踪成功目标数占视频序列目标总数的比例)为84.6%、ML(在整个序列中,有低于20%视频帧被跟踪成功目标数占视频序列目标总数的比例)为7.7%、Frag(视频序列目标真值所对应轨迹在跟踪中断开的次数)为9、IDS(在跟踪中,目标身份的改变次数)为4; 与其他同类型多目标跟踪算法相比,本文算法在FAF和Frag两个评估参数上表现出色。结论 融合特征能对目标进行较为全面的表达、判别性外观模型能有效地应用于跟踪片关联,本文算法能实现复杂场景下的多目标跟踪,且可以应用到一些高级算法的预处理中,如行为识别中的轨迹检索。

Objective A tracking-by-detection multi-target algorithm is proposed based on feature fusion and discriminative appearance models. Method To identify the responses of the detector in each sliding window, a dual-threshold strategy is adopted to perform low-level association and generate reliable tracklets between adjacent frames. Training samples are collected from these tracklets. Then, we merge several features to robustly describe the training samples and use the Adaboost algorithm to train the classifier, i.e., discriminative appearance model, online. Finally, the discriminative appearance model is used to link the tracklets into longer ones to form the final complete target trajectories by an iterative process. Result Experimental results on four challenging databases(TUD-Stadtmitte, TUD-Campus, TUD-Crossing, and Town-Center) show that the proposed method can efficiently deal with occlusions, target deformation, and background interference. The tracking results on the TUD-Crossing database are quantitatively analyzed, and the performance metrics of our algorithm are as follows: the FAF is 0.21, the MT is 84.6%, the ML is 7.7%, the Frag is 9, and the IDS is 4. The proposed method outperforms several state-of-the-art approaches in terms of FAF and Frag. Conclusion The multi-feature fusion is appropriate for target expression, and the discriminative appearance model is effective for tracklets association. The proposed algorithm exhibits satisfactory performance in a complex scene and can be further applied to the preprocessing of some advanced algorithms, such as trajectory retrieval in behavior recognition.