针对多目标跟踪算法中经常会面临的各种挑战, 如相机的突然运动、遮挡、误检和外观相似等情况, 提出一种基于核相关滤波(KCF)的分步关联框架。首先, 该算法采用基于卷积神经网络的目标检测器检测目标, 获得准确的检测结果。然后, 为了更好地预测目标的运动状态, 通过加权融合三种特征的跟踪结果, 为每个目标建立一个基于KCF算法的快速跟踪器。此外, 为了有效地降低碎片化轨迹的数量, 该算法通过跟踪片的置信度分步关联轨迹, 并在遮挡的情况下, 利用在线随机蕨重新检测目标。最后利用关联成功的检测信息自适应更新KCF算法中的尺度。实验结果表明, 与现有算法相比, 所提算法能够在各种复杂的条件下, 表现出强大和高效的跟踪性能。

A kernel correlation filter (KCF) with step-by-step association framework is proposed to aim at the various challenges, such as camera sudden movement, occlusion, false detection and appearance similarity in multi-target tracking algorithms. Firstly, the accurate detection results are obtained by utilizing a target detector based on the convolutional neural network. Then, a fast tracker based on the KCF algorithm is established for each target by weighted fusion of the tracking results of the three features to predict the motion state of the target. In addition, in order to effectively reduce the number of fragmented trajectories, this algorithm associates the trajectories step by step through the confidence of tracklets, and uses the online random fern to re-detect the target in the case of occlusion. Finally, the scale in the KCF algorithm is adaptively updated by using the associated successful detection information. Experimental results illustrate that the proposed algorithm displays powerful and efficient tracking performance under various complicated conditions compared with the existing excellent algorithms.