Localization of a mobile robot is a very important issue for robot's navigation. However, localization method with conventional wheel odometry has limits in case the wheel faces slippery conditions. As an alternative way, visual odometry has been researched continuously. However, this method alone has also difficulty for robust localization because wrong depth measurement can frequently occur and the error is accumulated continuously. Even though localization can be improved by using particle filter, this method is dependent on the accuracy of the reference map. For improving these drawbacks, this research utilized variable uncertainty useful for denoting accuracy of motion model from video information. Consequently, localization in the environments represented by inaccurate maps was improved compared to the conventional method.