A Method for Improvement of PDR method using Smartphones Considering Pedestrian Properties

Document Type : Original Article


1 School of Surveying and Geospatial Engineering, College of Engineering, University of Tehran

2 University of Tehran

3 Sahand University of Technology



Indoor positioning is one of the most challenging issues in location-based services. Smartphone-based Pedestrian Dead Reckoning (PDR) is commonly used as an indoor positioning system because it does not require infrastructure. The positioning estimation errors, however, are cumulatively increased over time. In this approach, step length estimation error is one of the main sources of positioning error. In this study, to improve indoor positioning accuracy using smartphone-embedded sensors, the pedestrian's gender and walking speed are considered effective factors in adjusting the parameters of step length estimation methods. Accordingly, collected data are divided into six classes based on walking speed (high, medium, low) and gender (female, male). K-Nearest Neighbors (KNN), Support Vector Machine (SVM), and Decision Trees (DTree) algorithms are adopted for classification. The classification accuracy of each of the KNN, SVM, and DTree algorithms are 93.4%, 92.4%, and 76.6% respectively. Moreover, the peak detection method is applied to identify the pedestrian's steps, and Weinberg and Ladetto methods are adopted to estimate step length. Step detection accuracy was 99.015%. Also, the error of step length estimations using Weinberg and Ladetto methods are 2.48% and 1.95%, respectively. In addition, the Extended Kalman Filter (EKF) filter is used for heading estimation, and fast walking results in the highest heading estimation error for both males and females. The mean and STD of the heading estimation error using EKF algorithms are 2.97 degrees and 2.99 degrees, respectively. In the final, the pedestrian's position is estimated according to the PDR method using estimated step lengths and estimated headings. Along a 25.8-meter path, using the Weinberg method for step length estimation, the average absolute and relative positioning errors are 0.76 and 2.95%, respectively. Moreover, using Ladetto s method for step length estimation, the average absolute and relative positioning errors are 0.92 and 3.57%, respectively.


Main Subjects