Nowadays smartphones are equipped with many sensors which applications can continuously invoke to acquire real-time sensor information, such as GPS tracking. Due to the resource-constrained nature of the smartphones, it is often beneficial if the processing of the sensor data is offloaded to a remote resource. However, the decision to offload the computation depends on a multitude of factors such as the hardware capabilities of the phone, the communication energy and latency and the characteristics of the stream computations, e.g., window size, sensor frequency and operational complexity.In this paper we introduce Kea, a profiling-based computation offloading system that automatically decides whether offloading is beneficial for smartphones. The decision making is based on two criteria: the power consumption of the application and the elapsed time for processing the sensor data. Our evaluation results show that unexpected factors such as CPU frequency scaling and the network state also influence the decision-making process. In addition, we show that Kea's profiling overhead is negligible.