Analysis of Output Domain More Reliable for Deep Learning-Based Step Length Estimation

Junu Park, Jae Hong Lee, and Chan Gook Park

Abstract: In this paper, we analyze the step length estimation (SLE) performance in deep learning (DL)-based pedestrian dead reckoning (PDR) regarding the input and output domain for achieving a high accuracy in SLE, thereby constructing a robust PDR. PDR consists of three parts: step detection, SLE, and heading estimation. In PDR, the step length is a crucial parameter that determines where the pedestrians are located and how much they have moved. Traditional methods for SLE include modeling methods based on the physical information and experimental experiences. The traditional methods are affected by heterogeneity which comes from a variety of the leg length, weight, walking speed, and motion of the pedestrian. DL-based PDR algorithms have been studied to overcome the weaknesses of the traditional methods. Due to the capability of DL to learn a nonlinear relationship between input and output, DL employs inputs at constant time intervals, in contrast to traditional methods that consider the relationship between output and step-size input. To examine reliable domains for achieving high accuracy in DL-based PDR, this paper analyzes the performance of DL-based PDR based on the input and output domain, with a specific focus on SLE. The input domain considered in this paper consists of time-aligned windows with constant time intervals and step-aligned windows generated by step events of pedestrian. Additionally, the output domain includes step length and walking speed of pedestrian, which are denoted distance-based and velocity-based model, respectively. The performance of each neural network model according to each domain is evaluated using a public dataset. The evaluation results demonstrate that the distance-based model, which takes step-aligned windows as input, performs better in terms of step length error and walking distance error. Furthermore, the results show that estimating step length directly rather than walking speed is a more robust approach for PDR.
Published in: Proceedings of the ION 2024 Pacific PNT Meeting
April 15 - 18, 2024
Hilton Waikiki Beach
Honolulu, Hawaii
Pages: 664 - 670
Cite this article: Park, Junu, Lee, Jae Hong, Park, Chan Gook, "Analysis of Output Domain More Reliable for Deep Learning-Based Step Length Estimation," Proceedings of the ION 2024 Pacific PNT Meeting, Honolulu, Hawaii, April 2024, pp. 664-670.
Full Paper: ION Members/Non-Members: 1 Download Credit
Sign In