The attitude measurement system (AMS) for unmanned surface vehicles (USVs) in GNSS-denied environments comprises dual micro-electro-mechanical system inertial measurement units (MEMS-IMUs), a Doppler Velocity Log (DVL), and a depth meter. One of the MEMS-IMUs is fixed to the ship, while the other rotates at intervals. By incorporating relative observations between the two MEMS-IMUs, the system's observability is significantly enhanced, leading to improved accuracy in attitude measurements. In dynamic environments, state transformation Kalman filter (STEKF) can obtain higher attitude measurement accuracy than Extended Kalman Filter (EKF). However, due to abrupt changes in underwater terrain, the velocity measurement noise of the DVL can increase, potentially affecting the accuracy of the STEKF. To address this challenge and achieve high-precision and stable attitude measurement, we propose a state transformation sequential adaptive filtering algorithm based on residuals (STAKF). This algorithm employs an adaptive method rooted in residual vectors to maintain the positive definiteness of the observation noise matrix during the recursive process, ensuring the stability of filtering results and superior adaptability to observation noise. Experimental results demonstrate that compared to STEKF, the proposed STAKF exhibits superior attitude measurement accuracy and enhanced adaptability to observation noise.