Online multi-object tracking (MOT) is an active research topic in the domain of computer vision. In this paper, a CRF-based framework is put forward to tackle the tracklet inactivation issues in online MOT problems. We apply the proposed framework to one of the state-of-the-art online MOT trackers, Tracktor++. The baseline algorithm for online MOT has the drawback of simple strategy on tracklet inactivation, which relies merely on tracking hypotheses' classification scores partitioned by using a fixed threshold. To overcome such a drawback, a discrete conditional random field (CRF) is developed to exploit the intra-frame relationship between tracking hypotheses. Separate sets of feature functions are designed for the unary and binary terms in the CRF so as to cope with various challenges in practical situations. The hypothesis filtering and dummy nodes techniques are employed to handle the problem of varying CRF nodes in the MOT context. In this paper, the inference of CRF is achieved by using the loopy belief propagation algorithm, and the parameters of the CRF are determined by utilizing the maximum likelihood estimation method. Experimental results demonstrate that the developed tracker with our CRF-based framework outperforms the baseline on the MOT16 and MOT17 datasets. The extensibility of the proposed method is further validated by an extensive experiment.