Road geometry information is an important information source in the autonomous driving perception system, which also plays an important role in the subsequent route planning. To realize the autonomous driving perception while the lane line is invisible and the signal of global positioning system is not available, a road geometry estimation based on the leading vehicle is proposed in this work. By modeling the relationship between the current vehicle, the preceding vehicle and the road, we can obtain the system motion model and the observation model. Then, the unscented Kalman filter framework is applied to filter the observed relative position, relative speed, and relative angel of the preceding vehicle and the angular velocity of the host vehicle, for estimating the curvature of current road. The experimental results on the simulation software car learning to act (Carla) showed that, in congested scenarios where lane line targets cannot be obtained and host vehicle cannot be accurately located, road geometry accuracy by the proposed method can be greatly improved in comparison with conventional map matching methods.