Smart camera networks raise challenging issues in many fields of research, including vision processing, communication protocols, distributed algorithms or power management. The ever increasing resolution of image sensors entails huge amounts of data, far exceeding the bandwidth of current networks and thus forcing smart camera nodes to process raw data into useful information. Consequently, on-board processing has become a key issue for the expansion of such networked systems. In this context, FPGA-based platforms, supporting massive, fine grain data parallelism, offer large opportunities. Besides, the concept of a middleware, providing services for networking, data transfer, dynamic loading or hardware abstraction, has emerged as a means of harnessing the hardware and software complexity of smart camera nodes. In this paper, we prospect the development of a new kind of smart cameras, wherein FPGAs provide high performance processing and general purpose processors support middleware services. In this approach, FPGA devices can be reconfigured at run-time through the network both from explicit user request and transparent middleware decision. An embedded real-time operating system is in charge of the communication layer, and thus can autonomously decide to use a part of the FPGA as an available processing resource. The classical programmability issue, a significant obstacle when dealing with FPGAs, is addressed by resorting to a domain specific high-level programming language (CAPH) for describing operations to be implemented on FPGAs.