Deep-learning Approach for Uncertainty Error Evaluation of Crowdsourced Trajectories and Navigation Database Generation

Ubiquitous indoor positioning technology plays an important role in providing indoor location-based services (iLBS) for the public. At this stage, crowdsourced multi-modal data fusion is regarded as an effective way to realize ubiquitous indoor positioning, especially for large-scale indoor spaces based on public daily-life trajectories and local positioning stations. Therefore, an effective uncertainty error evaluation method for daily-life trajectories is the key to generating a high-quality crowdsourced navigation database and further improving the performance of the final multi-source fusion system. To solve this problem, this paper proposes a deep-learning approach for autonomously evaluating the uncertainty error of crowdsourced daily-life trajectories, by learning and analyzing motion features extracted from pedestrian trajectories comprehensively from spatial and temporal perspectives. A novel deep-learning structure taking into account the spatiotemporal characteristics of the trajectory is modeled and related spatiotemporal features are extracted and modeled as the input vector of the proposed deep-learning structure. Real-world experimental results under generated trajectory datasets from large-scale indoor scenarios indicate that the proposed deep-learning structure can autonomously evaluate the uncertainty error of crowdsourced trajectories and realize much more accurate navigation database generation performance compared with existing state-of-the-art algorithms.