Abstract: The probability hypothesis density (PHD) filter is a novel approach to tracking an unknown and time-varying number of targets in presence of clutter, miss-detection. The Gaussian mixture PHD (GM-PHD) filter is an approximation implementation of the PHD filter. However the GM-PHD filter do not provide identities of individual target state estimates, which are needed to construct tracks of individual targets. Meanwhile, it is difficult for the GM-PHD filter to track the targets randomly appearing from uncertain position of the observation space because in order to generate target-birth density, the strict restricts on the appearing position of the birth targets exist in the GM-PHD filter. To solve the problems of the generation of the target-birth PHD and the track labeling in a nonlinear observation system, a new GM-PHD filter is proposed in which the measurements of sensors are used to generate the target-birth PHD, and the output state estimates of targets of the filter over time are associated with the additional identifying label in the output Gaussian components of the filter. The efficiency of the proposed algorithm is verified by simulations.