Sensor arrays configured in a distributed network offer the potential to accurately localize and track wideband targets. We use a maximum a posteriori penalty function (MAP-PF) tracking approach to derive the MAP estimate of the target state directly from the array data rather than from intermediate "measurements" of target bearings. A key feature of the approach is the use of the penalty function method of nonlinear programming to obtain a tractable solution. A sequential track state update procedure similar to the extended Kalman filter (EKF) is developed which updates the state first from the motion model, and then from the current array data. Penalized maximum likelihood estimates of the source bearing are computed for each array, which then act as a synthetic measurements in an EKF update of the state. The two-step estimation process is similar to traditional methods, except the processes are coupled via the penalty function. In the bearing estimation step, the current target state is used to guide the estimation process and help eliminate ambiguous or spurious estimates. In the track estimation step, estimated signal powers control the influence of the bearing estimates from each array on the final track estimate. The algorithm can be implemented in a decentralized manner where bearing estimation is performed at the arrays, and track estimation is performed at a central processing site. Simulation results for a relevant aeroacoustic ground scenario are presented.