Normal estimation is the basis for most applications using pointcloud, such as segmentation. However, it is still a challenging problem regarding computational complexity and observation noise. In this paper, we propose a normal estimation method for pointcloud using results from tensor voting. Comparing with other approaches, we show it has smaller estimation error. Moreover, by varying the voting kernel size, we find it is a flexible approach for structure extraction as well. The results show that the proposed method is robust to noisy observation and missing data points as well. We use a GPU based implementation of Sparse Tensor Voting, which enables realtime calculation.