In some incremental localization algorithms

error can be easily accumulated. Some centralized algorithms needs to collect information of the entire network

thus the communication cost is high. Aiming at these drawbacks

a semi-centralized localization algorithm based on support vector regression was presented. The base node collected the position of nodes and all connectivity information between anchor nodes as training samples to run the training procedure with support vector regression method. As a result

a regression function could be derived and was distributed to all sen-sors in the network. Then

normal nodes could perform the estimation of locations using the function. In order to increase the number of training samples

the normal nodes having minimum three anchor nodes as neighbors was located and became to anchor nodes with range based least-square method using RSSI. Analyses and simulation results show that the algorithm can reduce the overheads of communication and decrease the influence of ranging error

and has a high local-ization accuracy.