Distributed belief propagation is a promising technology for cooperative localization. Difficulties with belief propagation lie in achieving high accuracy without causing high communication overhead and computational complexity. In this paper, we propose an efficient cooperative localization algorithm based on distributed belief propagation and a new empirical indoor ranging error model, which can be applied to indoor localization systems with non-Gaussian ranging error distributions. To reduce the communication overhead and computational complexity, the algorithm passes approximate beliefs represented by Gaussian distributions between neighbours and uses an analytical approximation to compute peer-to-peer messages. The proposed algorithm is validated on an indoor localization system deployed with 28 nodes covering 8000 m², and is shown to outperform existing algorithms.