It has been shown that the dynamic environment around the mobile robot can be efficiently and robustly represented by the Bayesian occupancy filter (BOF) . In the BOF framework, the environment is decomposed into a grid based representation in which both the occupancy and the velocity distributions are estimated for each grid cell. In such a representation, concepts such as objects or tracks do not exist and the estimation is achieved at the cell level. However, the object-level representation is mandatory for applications needing high-level representations of obstacles and their motion. To achieve this, a natural approach is to perform clustering on the BOF output grid in order to extract objects. We present in this paper a novel clustering-tracking algorithm. The main idea is to use the prediction result of the tracking module as a form of feedback to the clustering module, which reduces drastically the complexity of the data association. Compared with the traditional joint probabilistic data association filter (JPDAF) approach, the proposed algorithm demands less computational costs, so as to be suitable for environments with large amount of dynamic objects. The experiment result on the real data shows the effectiveness of the algorithm.