Target localization is one of the interesting applications of sensor networks. Localization algorithms that use received signal strength (RSS) measurements at individual sensor nodes have been proposed. The maximum likelihood (ML) algorithm is known as a popular algorithm for target localization. In uniform propagation environments, theML algorithm has high accuracy to estimate a target location. Meanwhile, in nonuniform propagation environments, the ML algorithm has low accuracy, because this algorithm uses RSS from all the sensor nodes equivalently. The residual weighting (RWGH) algorithm has been proposed to reduce the effect of nonuniform propagation environments. This algorithm first sets subsets of sensor nodes. The target location is estimated using the sensor nodes in each subset. The final estimated result is the averaged value of the estimated results of all the subsets weighted by their reliabilities. The reliability of each subset is the residual error of the distances between the target and each sensor node estimated by two ways. In the RWGH algorithm, there may be a case that the reliability of each subset is high although the estimation error of the subset is large. This degrades the final estimated result. In this paper, we propose a localization algorithm to reduce the effect of the subsets that have large errors. The proposed algorithm tries to detect the area where the density of the estimated results of subsets is high and reflect this information to the final estimated result. In this algorithm, the sensor field is split into cells. Each subset votes its reliability for a cell that includes the estimated location with the subset and the reliability of the subset is added to the cumulative reliability of the voted subsets. The cumulative reliabilities of cells are used to calculate the final estimated result. We show that the proposed algorithm has higher localization accuracy than theML and RWGH algorithms by computer simulation.