In order to detect occlusion boundary in video sequences, a novel occlusion boundary detection approach based on unsupervised online learning is proposed. The occlusion related features of the frame to be detected in video sequences are extracted and the time length corresponding to the frame is calculated. The pixel points' occlusion related information in the frame to be detected is obtained using Hedge algorithm and combining time length with different occlusion features. The occlusion related information of different features is voted to accomplish occlusion boundary detection of current frame. The detection result of current frame is used to estimate the feature weight for next frame based on Online Boosting idea to realize the detection of subsequent frames. The proposed method changes the weight of different features through online learning idea to accomplish occlusion boundary detection, and does not need to obtain priori knowledge of video sequences in advance. Experimental results show that, compared with existing methods, the proposed method has higher accuracy and better generality.