Effective earthquake early warning (EEW) system is crucial in earthquake-frequent countries such as Japan. While itâ€™s critical to quickly send out alarms of potential large shakings, it is equally important to minimize occurrences of false warnings. False warnings typically result from incorrect estimates of event parameters (e.g. epicenter location, event origin time), which directly lead to erroneous estimate of event magnitude. In the following 50 days after the March 11th Tohoku earthquake, there have been over 70 warnings issued by Japan Meteorological Agency to the general public. Among those, 17 (25%) were false alarms of actual detected intensity less than 2. The reason behind the abnormally high false alarm rate was inherent in the current EEW design. The current EEW system canâ€™t handle detection and parameter estimation of multiple simultaneous events, which are common during the high seismic activity period after a major earthquake. In the scenario when two or more small events occur about the same time but distances apart, the system generates erroneous estimates based on the incorrect assumption that there is only one event. One possible solution to this problem is to exploit the spatial and temporal correlations of sensor readings. Seismic waves travel in a certain manner that has been well characterized by seismologists. Combining the physical model of seismic wave progression and a probabilistic model of the sensor measurements, the problem naturally falls into the realm of Bayesian parameter estimation and multi-target tracking - a technique that has been studied extensively in the field of control and tracking. Exact Bayesian estimation with many parameters is very computationally intensive, so we implemented such technique through sequential Monte Carlo (also known as particle filter). To illustrate this better, consider two stations A and B that both detect abnormally large shaking. Given the speed of wave, the distance between A and B, and existing events, one can compute the probability that both stations detect the same event. Furthermore, this probabilistic approach can be integrated into the current EEW system to improve estimation accuracy. The algorithm is tested with real seismic data recorded by the ~200 JMA stations in Japan. In an effort to estimate the parameters of the March 11th Tohoku event, the algorithm shows encouraging fast convergence rate on the 5 parameters we want to estimate – latitude, longitude, depth of the epicenter, magnitude and starting time of the event. It was also tested against a 5-minute segment of data, during which the EEW system gave out false warnings, and is able to identify and separate out 3 independent events occurred during that period with relatively accurate estimates of epicenter and magnitude.