Automatically detecting and tracking the motion of Myxococcus xanthus bacteria provide essential information for studying bacterial cell motility mechanisms and collective behaviors. However, this problem is difficult due to the low contrast of microscopy images, cell clustering and colliding behaviors, etc. To overcome these difficulties, our approach starts with a level set based pre-segmentation of cell clusters, followed by an enhancement of the rod-like cell features and detection of individual bacterium within each cluster. A novel method based on "spikes" of the outer medial axis is applied to divide touching (colliding) cells. The tracking of cell motion is accomplished by a non-crossing bipartite graph matching scheme that matches not only individual cells but also the neighboring structures around each cell. Our approach was evaluated on image sequences of moving M. xanthus bacteria close to the edge of their swarms, achieving high accuracy on the test data sets.