In this paper, we propose a method to directly estimate a camera's motion parameters by using normal flow vectors. In contrast to traditional methods, which tackle the problem by calculating optical flows or establishing motion correspondences, our proposed approach does not require conventional assumptions about the captured scene, such as consistent smoothness or distinct feature availability. In the proposed algorithm, the normal flows are classified into different groups, and each group will provide a possible solution regarding the camera's motion parameters. Then, the strategy of hypothesis and confirmation is adopted to eliminate the incorrect solutions. Finally, the optimal solution is obtained via the clustering algorithm. We have tested the proposed method on both synthetic image data and real image sequences. The experimental results illustrate the feasibility and reliability of the algorithm.