Accurate indoor localization is difficult because building structures attenuate satellite signals, making GPS unreliable indoors. WiFi fingerprinting offers an alternative that uses existing network infrastructure: a device’s measurements are compared with fingerprints recorded at known reference points. This thesis evaluates WiFi Received Signal Strength (RSS), Round-Trip Time (RTT), and their combination for estimating coordinates within a floor and for identifying the floor. RTT ranging measurements are used as fingerprint features rather than for trilateration, so position estimation requires reference-point coordinates but not access point (AP) coordinates.

Experiments were conducted on three floors of the Brendan Iribe Center for Computer Science and Engineering at the University of Maryland, using 297 training reference points and 61 independent test locations. RTT measurements were widely available on the fifth floor, partially available on the fourth floor, and insufficient for localization on the third floor. Six methods were evaluated: KNN and probabilistic HORUS using RSS and RTT separately; Weighted KNN RSS+RTT, which combines normalized fingerprint distances; and Extended HORUS RSS+RTT, which combines Gaussian log-likelihoods with regularized RTT variance estimates. Four AP-matching strategies were compared: All APs, AP Intersection, APs Observed at Test, and Jaccard Similarity. Parameters were selected by LOOCV and evaluated on the independent test locations. Graph Fréchet coordinate estimation and the effect of device orientation on WiFi measurements were also examined.

On the fifth floor, Weighted KNN RSS+RTT achieved the lowest mean independent-test error with its LOOCV-selected configuration (2.25 ft), followed by HORUS RTT (2.32 ft) and KNN RTT (2.35 ft). Extended HORUS RSS+RTT (2.69 ft) improved on the RSS-only methods but not on the RTT-only methods. On the fourth floor, only 14 of the 21 test locations had valid RTT measurements, which limited the coverage of the RTT and fusion methods. On the third floor, RSS-only localization reached a lowest mean test error of 12.82 ft. The AP-matching strategy strongly affected both accuracy and the number of locations that could be localized. Default values for missing measurements preserved coverage but increased the errors of several RTT and fusion configurations, whereas requiring shared APs reduced coverage where RTT was sparse. Graph Fréchet estimation reduced the error of some configurations, for example KNN RSS on the fifth floor from 3.41 ft to 2.87 ft, but not consistently. For floor classification, KNN RSS and Weighted KNN RSS+RTT correctly classified 60 of the 61 test locations (98.36%), while RTT-only methods could evaluate only 30 locations.

These results show that RTT can provide accurate coordinate estimates where sufficient ranging measurements are available, while RSS provides broader coverage and more reliable floor classification. The benefit of combining RSS and RTT depends on the AP matching strategy, the fusion weights and measurement availability.