During transcription, RNA polymerase competes for space on the DNA with other DNA binding proteins and higher order DNA structures acting as roadblocks. Though it is known that individual polymerases often slow down when forcing roadblocks, the effect of crowding on transcription as a whole is not clear. Based on quantitative theoretical modeling, we show that interactions with roadblocks induce a strong kinetic attraction between polymerases, causing them to self-organize into pelotons. Peloton formation explains observed nucleosome and polymerase density profiles close to the initiation site on highly transcribed genes, and how these densities depend on induction levels. At termination, pelotons translate into transcriptional bursts that dispaly the same characteristics as those observed in vivo. Our model thus unifies common spatial and temporal transcription patterns as arising from a non-specific interaction between roadblocks and polymerases. The generality of our model suggests that peloton formation might be ubiquitous in systems where molecular motors interact with dynamic roadblocks.