Recently, low-cost acceleration sensors have been widely used to detect earthquakes due to the significant development of MEMS technologies. It, however, still requires a high-density network to fully harness the low-cost sensors, especially for real-time earthquake detection. The design of a high-performance and scalable networked system thus becomes essential to be able to process a large amount of sensor data from hundreds to thousands of the sensors. An efficient and accurate earthquake-detection algorithm is also necessary to distinguish earthquake waveforms from various kinds of non-earthquake ones within the huge data in real time. In this paper, we present CrowdQuake, a networked system based on low-cost acceleration sensors, which monitors ground motions and detects earthquakes, by developing a convolutional-recurrent neural network model. This model ensures high detection performance while maintaining false alarms at a negligible level. We also provide detailed case studies on two of a few small earthquakes that have been detected by CrowdQuake during its last one-year operation.