Abstract: In a GPS-denied environment, visual odometry based on sequential stereo images can be applied to correction of accumulated localization errors from dead-reckoning. Visual odometry however, preforms. poorly or even fails under poor texture or illumination conditions. To achieve high precision and robust localization, a federated filter based localization method is proposed that integrates stereo cameras, the IMU, and wheel odometer. The two sub filters in the federated filter, constructed by a IMU integrated with the wheel odometer and stereo cameras provide robust localization when visual odometry fails. Experimental results in an underground field demonstrate robust performance and improvement in localization accuracy.