The maximum-likelihood probabilistic multi-hypothesis tracker (ML-PMHT) is a tracking method whose flexibility and scalability derive from relinquishing the assumption that each target emits at most one “hit” per scan of the sensor. This is an ML method that essentially reduces to an optimization problem—recursively maximizing a likelihood function that is simple to evaluate given a batch of observations. Unlike maximum a posteriori or minimum mean squared error (MMSE) trackers, this likelihood maximization tracker requires neither prior knowledge about target motion nor measurement association, making it conceptually easy to work with. Here, this method is used to track targets in a three-dimensional “global” space with observations provided by multiple two-dimensional sensors placed through-out the global space. Since the observation model is non-linear, the likelihood maximization is done via hill climbing. For this purpose, we also address the issue of “hill finding.” Due to the presence of clutter in the measurement model, the likelihood is a multi-modal function of the parameter space. That is, there are multiple hills in the likelihood function, and it is of great advantage to the tracker to initialize the hill climber close to the right hill—the one whose peak is the global maximum. In this work, we present a data-driven method of initializing the hill climber based on the received observations.