This work proposes an extension of Bundle Adjustment to dynamic scenes. In the setting of one or multiple cameras moving in a dynamic environment, the camera pose and the 3D positions of static and moving objects are reconstructed from the captured image sequences. An efficient, low- dimensional representation of the scene is introduced, which is based on approximating trajectories by linear combinations of trajectory bases. Our reconstruction approach requires no knowledge about the objects, not even which are moving or static and is, in difference to other approaches, able to deal with incomplete and noisy data. Experimental evaluation in simulation as well as with real data shows its effectiveness in reconstructing dynamic scenes from moving cameras.