State estimation for legged locomotion over a dynamic rigid surface (DRS), which is a rigid surface moving in the world frame (e.g., ships, aircraft, and trains), remains an under-explored problem. This paper introduces an invariant extended Kalman filter that estimates the robot's pose and velocity during DRS locomotion by using common sensors of legged robots (e.g., inertial measurement units (IMU), joint encoders, and RDB-D camera). A key feature of the filter lies in that it explicitly addresses the nonstationary surface-foot contact point and the hybrid robot behaviors. Another key feature is that, in the absence of IMU biases, the filter satisfies the attractive group affine and invariant observation conditions, and is thus provably convergent for the deterministic continuous phases. The observability analysis is performed to reveal the effects of DRS movement on the state observability, and the convergence property of the hybrid, deterministic filter system is examined for the observable state variables. Experiments of a Digit humanoid robot walking on a pitching treadmill validate the effectiveness of the proposed filter under sensor noise and biases as well as large estimation errors and DRS movement.