Abstract: Multi-Bernoulli (MB) filter based on stochastic finite set theory has been demonstrated as a promising algorithm for tracking multiple targets with unknown and time-varying number of objects. However, for visual multi-object tracking (VMOT) in a complex scenario, it is difficult to solve the interference between close targets and background clutter. Especially, when the targets are close to each other and occluded, it will lead to the decrease of tracking accuracy and even target missed tracking. To solve these problems, under the framework of multi-Bernoulli (MB) filter, deeply analyzing the feature information of the target and introducing the anti-interference convolution feature, this paper proposes an effective VMOT algorithm by integrating convolution features into the framework of MB filter. Moreover, in the process of object state extraction, the adaptive scheme of template update is proposed by using an adaptive learning rate, which makes the template adapt to the scale variation and handle the problem of closely-spaced object tracking effectively. Finally, the particle labelling technique is employed to realize visual multi-object tracking. Experimental results show that, the proposed algorithm can effectively distinguish the closely-spaced object in a complex scenario, and has a high tracking accuracy.

Key words: multi-Bernoulli (MB) filter, convolution feature, adaptive learning, visual multi-object tracking

摘要： 基于随机有限集理论的多伯努利滤波方法能够有效处理多目标跟踪中数目未知且时变的问题，但难以适应复杂环境下视频多目标跟踪中目标之间或背景等干扰问题，尤其是目标相互紧邻和被遮挡时，会导致跟踪精度下降，甚至目标漏跟。针对该问题，在多伯努利滤波框架下，深度分析目标的特征信息，引入抗干扰的卷积特征，提出基于卷积特征的多伯努利视频多目标跟踪算法，并在目标状态提取过程中，进一步提出模板更新，使用自适应学习速率进行更新，适应目标的变化，以解决目标紧邻相互干扰的问题。最后，引入粒子标记技术，实现对视频多目标的航迹跟踪。实验结果表明，提出算法能够有效区分复杂环境下的紧邻多目标，且具有较好的跟踪精度。

关键词: 多伯努利滤波, 卷积特征, 自适应学习, 视频多目标跟踪