The active surveillance of manoeuvring targets with multiple dynamic cameras requires an effective planning methodology to dynamically select and position the groups of cameras for optimal performance. This paper presents such a methodology for the real-time reconfiguration of a multicamera active-vision system for the recognition of dynamic objects in the presence of multiple static and/or mobile obstacles.An agent-based coordination strategy determines how many and, specifically, which cameras should be used at each data-acquisition instant in order to optimize the performance of the surveillance system. A positioning strategy determines the optimal location of each chosen camera at all dataacquisition instants. The proposed sensing-system reconfiguration methodology has been implemented on an experimental prototype set-up for automated object recognition via Shapefrom- Shading (SFS). In contrast to previously proposed algorithms, the recognition method presented in this paper does not require images to be taken at constant viewing angles. Furthermore, our algorithm fuses data from several object images acquired from varying viewpoints and with different lighting conditions. Our simulations and experiments have shown that multi-camera active sensing and data fusion tangibly increase the accuracy and robustness of a recognition process.