Stride detection for pedestrian trajectory reconstruction: A machine learning approach based on geometric patterns

Beaufils, Bertrand; Chazal, Frederic; Grelet, Marc; Michel, Bertrand · 2017 · Crossref

DOI: 10.1109/ipin.2017.8115867

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of accurate pedestrian trajectory reconstruction using inertial sensors worn on the ankle, specifically focusing on robust stride detection. While Global Positioning System (GPS) navigation is common, it fails in indoor or obstructed environments. Existing inertial navigation systems often rely on Zero Velocity Update (ZUPT) techniques to limit error accumulation, but these require precise detection of stride phases. Current methods struggle with atypical movements—such as small steps, side steps, or backward walking—and frequently produce false positives when the wearer is stationary but moving their ankle, such as while sitting. The authors propose a machine learning approach based on geometric patterns to detect strides accurately across diverse activities and critical non-walking situations. The methodology involves a four-step algorithm. First, candidate stride intervals are identified using peaks in acceleration norm and local minima in angular velocity norm. Second, sensor orientation is normalized by computing a rotation matrix that aligns the gyroscope data with a defined body frame, using 3D reference patterns for seven activities (including walking, running, and stair climbing). Third, the "forward swing" phase is modeled by fitting the gyroscope’s y-axis data to 1D reference patterns derived from empirical data for six specific activities. Finally, a binary classification is performed using a Gradient Boosting Tree (GBT) algorithm. This classifier utilizes 2,695 features derived from frequency domain analysis, alignment residuals, and swing pattern comparisons to distinguish true strides from non-stride movements. The model was trained on a dataset of approximately 12,000 intervals (6,000 positive and 6,000 negative), validated via 10-fold cross-validation. Experimental results demonstrate the algorithm’s effectiveness in two scenarios. In a controlled test involving nine loops of varied movements—including running, side-stepping, backward walking, and spinning—the system maintained a trajectory error of less than 30 cm over 110 meters, with no missed strides confirmed by video. In a real-world test involving an office worker over 5.5 hours, the system correctly handled natural walking, stair climbing, and periods of sitting with ankle movement. The trajectory error at three verified waypoints was under 50 cm, and crucially, no false strides were detected during stationary periods. The significance of this work lies in its ability to detect atypical strides and reject false positives in non-walking contexts, addressing key limitations of existing literature. By combining geometric pattern modeling with GBT classification, the method enables reliable ZUPT-based trajectory reconstruction in complex, real-life environments. The authors conclude that while stride detection is robust, further improvements in estimating ankle speed during the stance phase could enhance overall trajectory accuracy.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• pedestrian behavior perception