Localization of the sensor nodes is a major obstacle for practical applications of video sensor networks. This paper present a novel localization technique based on vision in wireless sensor networks. On the assumption that sensor nodes can be recognized in an image, a sensor node firstly direct its Field-of-View (FoV) to an anchor or localized node, then we can get the orientation of anchor node relate to the sensor node to be localized. If two or more anchors can be found in sensing area, a series of equations will be obtained. They can be solved using minimum mean square error rule, and the solution of the overdetermined equations mentioned above is the estimation of the node position. The experiments indicate that a promising performance can be achieved in determining the exact node location using a small number of anchors.