Human motion change detection is a challenging taskfor a surveillance sensor system. Major challenges includecomplex scenes with a large amount of targets and confusors,and complex motion behaviors of different human objects.Human motion change detection and understandinghave been intensively studied over the past decades. In thispaper, we present a Hierarchical Gaussian Process DynamicalModel (HGPDM) integrated with particle filter trackerfor humanmotion change detection. Firstly, the high dimensionalhuman motion trajectory training data is projected tothe low dimensional latent space with a two-layer hierarchy.The latent space at the leaf node in bottom layer representsa typical humanmotion trajectory, while the root node in theupper layer controls the interaction and switching amongleaf nodes. The trained HGPDM will then be used to classifytest object trajectories which are captured by the particlefilter tracker. If the motion trajectory is different fromthe motion in the previous frame, the root node will transferthe motion trajectory to the corresponding leaf node. Inaddition, HGPDM can be used to predict the next motionstate, and provide Gaussian process dynamical samples forthe particle filter framework. The experiment results indicatethat our framework can accurately track and detect thehuman motion changes despite of complex motion and occlusion.In addition, the sampling in the hierarchical latentspace has greatly improved the efficiency of the particle filterframework.