With the booming of artificial intelligence and the advance of embedded systems, smart surveillance systems can now deploy video analysis techniques, such as human detection and object classification, on surveillance cameras directly. Such an approach decentralizes the computation and allows elastic camera deployment without upgrading the centralized server. Nevertheless, taking human detection as an example, cameras at different locations could have very different numbers of people walking by at different times. This condition could make specific cameras run their detection tasks at the highest frame per second (FPS) but still lose information if FPS is lower than a certain value, while a small group of cameras has almost zero tasks, thus wasting valuable computation power. In addition, as the smart surveillance system grows, cameras deployed at different periods could have different specifications and computation capabilities; therefore, properly utilizing all the computation resources has become essential for smart surveillance systems. The aforementioned observations motivate this study to propose a novel computation souring method by following the concept of crowdsourcing to enable efficient resource sharing within smart surveillance systems and maximize the amount of retrieved information. Promising results have been demonstrated through a series of experiments with different smart cameras of different specifications.