The importance of pedestrian detection in many applications has led to the development of many algorithms. In this paper, we address the problem of combining the outputs of several detectors. A pre-trained pedestrian detector is seen as a black box returning a set of bounding boxes with associated scores. A calibration step is first conducted to transform those scores into a probability measure. The bounding boxes are then grouped into clusters and their scores are combined. Different combination strategies using the theory of belief functions are proposed and compared to probabilistic ones. A combination rule based on triangular norms is used to deal with dependencies among detectors. More than 30 state-of-the-art detectors were combined and tested on the Caltech Pedestrian Detection Benchmark. The best combination strategy outperforms the currently best performing detector by 9% in terms of log-average miss rate.