This research aims to develop a method of estimating road boundaries using a 3D LIDAR. Our previous methods estimate road boundaries on a flat terrain with multiple sensory feature and particle filter. However, they are not suitable for unpaved roads on a complicated terrain such as mountain roads. In this research, we extend the road model used in these methods to three dimensions and change the state transition method accordingly. We tested the proposeed method using a simulated road including a gradient change and confirmed that the road boundaries can properly be estimated with respect to the gradient change.