Week 4 Commentary

Low Cost Vision-Aided IMU for Pedestrian Navigation

Smartphones have a lot of sensors that can be used when GPS is not available. The issue is those sensors are not those accurate and then errors of measurement increase rapidly. One solution presented in this paper is based on computer vision. There is no need for additional structure.
The idea of the solution is to use the camera of the device. For each frame, around 300 interest points are selected. Then consecutive images are matched using those points. The difference between the two images gives a rotation and a translation matrix. Result is accurate with an unknown scalar factor. One way to find this factor is buy using the height of the device. Because the computation of the rotation is less accurate than the internal measurement of sensor, only the translation is used to correct the measurement.

The test protocol starts by 97 seconds of standing and walking positions with GPS measurement. This sequence allows computing the eight of the phone and estimates the biases of sensors. Then, the GPS is disconnected. After 60 seconds, without computer vision correction, the horizontal error more than 200 meters. With computer vision correction, the error is 3 meters after 6 minutes. Two issues can be noticied. The solution is slightly shorter than the trajectory. This is because the system does not have the real height of the phone. Furthermore, the yaw error of the INS is not corrected because the computer vision only provides translation feedbacks.

This experiment is not real-time based. The trajectory is computed using Matlab because computer vision is not cheap. On way to improve this point is to use INS measure to prematch frames. Then the search area to match points between them can be reduced.

2D/3D Indoor Navigation Based on Multi-sensor Assisted Pedestrian Navigation in Wi-Fi Environments

GPS positioning does not provide enough accuracy to be used in indoor environment. Initial measurement unit based on pedestrian dead reckoning works only for a short time. One way to improve the system is based on Wifi. The system can be decomposed on two parts. In a first time, a database, or a radio map has to be created. The signal strength of each access point of the network is computed on points with known position. When the map is created, the localization can be processed. Received signals strength are measured and compared with known point using the least square methods. After computation, approximation of distance to each Access Point of the Wifi Network is known. The barometer is used to get the height of the user. Combining those two techniques allows three-dimensional tracking.

The experimentation is made on a two floors university building of 40 meters square. Test with just WIfi positioning gives a mean error of 3.23. When the system combining Wifi, barometers and pedestrian is used, the error is either 2.11 or 1.65 according of the kind of Kalman filter used. The system is viable and can be used for providing accurate indoor localization. This study also shows that for Wifi positioning systems, there is not a real need to adapt the Kalman filters.

Comments are closed.