Swarm localization, cooperative robot localization in swarm robotics, has a significant role in a swarm robot system and requires much deliberation for its estimation scheme. As such, designing stochastic hidden Markov model, in a way a variety of conditionally dependent, observed random variables such as measurements are effectively chosen and properly integrated into the probability distribution of a belief, is very important. In this paper, we propose swarm EKF localization, a hybrid of two inference algorithms, extended Kalman filter (EKF) and belief propagation (BP), with a capability of choosing how many dependencies of random variables are exploited in inference using the concept of neighborhood. Also, this paper presents a numerical experiment result of swarm EKF localizations. In conclusion, we could confirm that 2nd order neighborhood EKF has an overall better estimation performance compared to conventional 1st order neighborhood EKFs.