Bipedal robots have seen significant interest from academia and industry for their potential to efficiently traverse unstructured environments, such as disaster zones, industrial infrastructure, and cluttered homes. Autonomous bipedal walking in the wild remains an unsolved challenge, however, partially due to the difficulty of perceiving and reacting to obstacles in real-time, while maintaining balance. Existing walking control approaches for rough terrain decompose the problem into separate footstep planning and motion control tasks, limiting their usefulness for dynamic, underactuated robots. We first present a hierarchical walking controller which can generate dynamic walking motions over stepping-stone like terrains by jointly optimizing over foothold choices and motion plans. This controller is formulated as a Mixed-Integer Quadratic Program, and can be solved at 50-200Hz, depending on the complexity of the terrain. We then propose a perception system for generating a convex polygon representation of the terrain in real-time from robot-mounted depth cameras. The controller and perception system are demonstrated on the underactuated bipedal robot Cassie.