In this paper, we propose a training-free method for moving object detection in video sequences. Our method is mainly based on a novel clustering algorithm of accu-racy and simplicity. For each frame, dense optical flow between its previous frame and itself is firstly measured. Then for each region whose optical flow is high, the clustering method is applied on the histogram of optical flow orientation to segment different moving objects which are close to each other. Lastly, the consistency of motion vectors of each moving object candidate is veri-fied and the final detecting results are obtained. Experiments on videos in three public datasets show that our algorithm achieves a fast speed of at least 8.01 frames (compared to 1.25) per second and a high recall of at least 87.2% (compared to 83.5%) while the preci-sion is 93.5% (compared to 89.8), which outperform the state-of-art algorithm.