The accelerometer outputs the second derivative of the position, so it is theoretically feasible to obtain the position by double integrating the accelerometer data. However our experiments showed that the data outputted by this sensor is too noisy to get an estimation of the position. Figure 2 shows the results of an experiment comparing the ground truth 2D trajectory and the trajectory estimated with the accelerometer data, while the user walked holding the phone upright. The graph shows a bird’s eye view, with an equal number of points in both curves. An accurate trajectory estimate would overlap the rectangular ground truth; in contrast, the accelerometer-based position estimate was wildly off. Another solution is to use the GPS data which gives the location of the user with a few meters error. Depending on the application, that can be adequate. However, if the system is used indoor there is usually no GPS signal avail- able, so the position cannot be estimated with the cellphone sensors. Therefore in our tests, if there is no GPS signal available, we initialize the user location with the last computed position (i.e., the position computed for the previous frame). Where a GPS signal is not available, we assume the user does not walk more than two meters between two pictures, so that we can initialize his location to the previous computed location.