Effective Vehicle Tracking in Dynamic Indoor Parking Area based on Hidden Markov Model and Online Learning

As a fundamental requirement for Intelligent Transportation System (ITS), reliable and pervasive vehicular localization has attracted considerable attention. However, it is still challenging to enable such services in dynamic indoor parking environments where Global Navigation Satellite System (GNSS) is not available. In view of this, this work first proposes a Hidden Markov Fusion (HMF) algorithm to fuse WiFi fingerprinting localization with inertial sensors based Dead Reckoning (DR), in which the WiFi fingerprinting localization result is modelled as the emission probability and the displacement inferred by DR is modelled as the transition probability. On this basis, a forward probability is derived to estimate the distribution of a vehicle's position. Moreover, considering that signal features may vary over time in dynamic indoor parking environments, we further propose an online learning framework, which contains an online evaluation method to assess the accuracy of WiFi fingerprinting localization results and an modified Homogeneous Online Transfer Learning (HomOTL) algorithm to continuously update the fingerprinting localization model. Finally, we implement the system prototype and give comprehensive performance evaluation in realistic indoor parking environments, which conclusively demonstrates the effectiveness of the proposed solutions.