This paper proposes a general methodology to model pedestrian destination choice from WiFi localization in multi-modal transport facilities (e.g., airports, railway stations). It is based on the output of Danalet et al.(2014) method to generate candidates of activity episode sequences from WiFi measurements, locations of activities on a map and prior information. Destination choice is nested to the activity choice. An individual first chooses an activity (Danalet and Bierlaire, 2015), and then selects the destination where to perform it. We propose an approach to model destination choice accounting for panel nature of data. We compare static, dynamic strictly exogenous and dynamic with two different agent effect corrections models with inspiration from Wooldridge (2002) method. In a case study using WiFi traces on EPFL campus, we focus on one activity: catering. The choice set contains 21 alternatives on campus (restaurants, self-services, cafeterias, ...). Our models reveal that the choice of a catering facility especially depends on habits (e.g., where an individual ate the previous time), distance to walk from the previous activity episode (calculated with a weighted shortest path algorithm) and destination specific determinants. Price has a non-significant impact in this case study, most likely because the price range on campus is narrow. The models are successfully validated using the same WiFi dataset.