To overcome the difficulty of dense sampling in a high-dimensional hand state space, an improved Particle Swarm Optimized Particle Filter (PSOPF) algorithm was proposed to track articulated hand motion from single depth images obtained by a Kinect sensor. Firstly , a 3D hand model was built with basic geometric primitives, and the nodes with degrees of freedom (DOFs) were added into the model to generate the hand pose hypotheses to compare with the observation feature in the tracking process. Then, the single depth images were used as the only input,and the system observation model was established by combining depth features and silhouette features in the framework of particle filter. Finally, the PSOPF was applied to articulated hand tracking. To avoid the premature convergence in the high-dimensional space, the global search ability of the algorithm was improved by applying simulated annealing and partial randomization on the particles. Experiments were conducted both on synthetic data and real sequences for evaluation of the proposed method. It shows that the average of the joint angle errors for proposed method is about 2.3° and the standard deviation is about 1.7°, which are better than those of the standard Particle Filter(PF) and the standard Particle Swarm Optimization (PSO) method. The results show that the proposed method may track 3D articulated hand motion accurately and robustly from single depth images.