Target localization is an important application in sensor networks. Although the localization problem in acoustic /radar sensor networks has been intensively studied, few consider the problem in camera sensor networks. In this paper, we propose a method to utilize the active cameras to localize the target collaboratively. We describe the perspective projection of target firstly, and then utilize the Gaussian error model to establish a more practical target projection model. On the basis of the target projection model, we present a vision-based localization model, and use Shannon entropy to quantify the localization gain due to the observation of camera. Case study demonstrates the validity and efficiency of our design.