Numerous applications require accurate personal navigation for environments where neither GPS signals nor infrastructure beacons, such as WiFi, are available. Inertial navigation using low-cost sensors suffers from the noisy readings which leads to drifting errors over time. In this paper, we introduce a novel inertial navigation approach ProbIN using Bayesian probabilistic framework. ProbIN models the inertial navigation problem as a noise channel problem where we want to recover the actual motion/displacement of the user from the noisy sensor readings. Building on the top of dead reckoning, ProbIN learns a statistical model to map the noisy sensor readings to user's displacements instead of using the double integral of the acceleration. ProbIN also builds a statistical model to estimate the a priori probability of a user's trajectory pattern. Combining the mapping model and the trajectory model in a Bayesian framework, ProbIN searches for a trajectory that has the highest probability given the sensor input. Our experiments show that ProbIN significantly reduces the error of inertial navigation using low-cost MEMS sensors in mobile phones.