To increase the immersion for brain-computer interface (BCI)-visual scene system, a brain-controlled virtual reality (VR) roaming system in museum scene is designed in this paper using MI-based BCI and virtual reality engine. Firstly, an online logarithmic band power (LBP) feature extraction algorithm combined with support vector machine (SVM) classifier is applied to recognize the left / right motor intentions. Secondly, a real-time communication mechanism between BCI platform and virtual scene is established to real-time control the virtual human in the virtual scene based on the recognition results. Then, a newly designed VR museum scene is accomplished by Unity engine to make subjects feel more relaxed and immersed. Finally, experimental results reveal that the motor imagery tasks of left / right hand are distinguished online effectively, and the virtual reality roaming scene is driven accurately in real time by the proposed system.