Localization, estimating the positions and orientations of a set of camera nodes, is a critical first step in camera-based sensor network applications such as geometric estimation and scene reconstruction. We propose a distributed algorithm for camera network localization based on feature point correspondences between multiple cameras with sparse overlapping view structure. We prove convergence of the iterative piecewise-linearized algorithm using the projection onto convex sets (POCS) principle, since this algorithm corresponds to projecting onto subspaces that approximately overlap. We provide bounds on convergence rates and worst-case errors and show experimental results from actual images. Finally, we introduce a new technique to obtain initial localization estimates based on trajectory observations.