Multi-Bernoulli (MB) filter has been demonstrated as a promising algorithm for tracking multiple point targets with unknown and time-varying number of targets. However, for the visual multi-object tracking (VMOT), the tracking accuracy will decrease severely due to the problems of closely-spaced object, occlusion and scale variation. To solve these problems, an adaptive VMOT algorithm based on the framework of the MB filter is proposed in this paper. First, the region feature covariance (RFC) is employed to enhance the ability of anti-interference, and then the discrimination strategy for the closely-spaced objects and the adaptive estimate strategy for scale variation are developed according to the constraints of tracking boxes. Finally, the particle labeling technique is introduced to identify the track of each target. Experimental results validate superior performance of the proposed algorithm over the traditional MB-based VMOT algorithm in scenarios that include occlusion, background clutter, scale changes, and closely-spaced objects. © 2019, ICIC International. All rights reserved.