In tele -present or tele -operation applications, an inte l- ligent observer is a virtual camera that can perform autonomous motions to assist a user in accomplishing de- sired observing tasks. In this paper we consider the problem of generating and modifying the motions of an intelligent observer tracking a moving target in an online manner. Given the target's trajectory, we first plan the observer's motion to maintain visibility with the target while optimizing certain time-dependent camera-specific criteria. During the execution of the motion, we allow a user to interactively command the observer to deviate from the planned path while guaranteeing visibility with the target in the remaining motion. The online nature of the problem imposes a co n- straint on the planning time allowed in each execution cycle. Efficiency of the motion planner becomes the greatest con- cern. In this work we formulate the problem in the relative configuration-time space (CT-space) of the observer and adopt a best-first algorithm to search for a feasible path. In addition, we propose an incremental backward search algo- rithm in the CT-space to handle dynamic modification of the planned path. Simulation experiments show that the planner can generate a feasible tracking path in fractions of a sec- ond and maintain validity of the planned path for each exe- cution cycle in tens of milliseconds. This planning efficiency is achieved partly due to an efficient visibility detection al- gorithm that we have developed for planning tracking mo- tions. This implemented intelligent system is part of an in- teractive3D auto-navigation system for virtual factories. We believe that these experimental results can inspire a new line of potential applications that could benefit from active, in- telligent observations.