Accurate Step Counting and Real-Time Step Length Estimation Based on Pedestrian Single-Step Acceleration Area Features

This project is a 2026 entry in the Scientific Paper Track of the National College Students Surveying and Mapping Innovation and Entrepreneurship Intelligence Competition. It addresses two critical inputs to smartphone pedestrian dead reckoning (PDR): step detection and step-length estimation. The method targets missed and false detections during variable-speed walking, start-stop transitions, individual differences, and sensor-placement changes, as well as the limited adaptability of empirical step-length models.

Using raw tri-axial resultant acceleration, the method treats single-step acceleration area as a unified descriptor of gait-impact intensity and duration, applying the same feature to candidate-gait correction, accurate step counting, and real-time step-length estimation.

The step-counting pipeline first constrains peak-valley candidate windows using the upper cadence limit of fast walking. A 0.2g dynamic-acceleration threshold, valley aggregation, and crossed peak-valley correction reduce duplicate candidates caused by non-walking disturbances and local multi-peak responses.

Positive and negative single-step acceleration areas are computed within each candidate gait segment. Neighboring area features are then merged to correct over-segmentation, small-amplitude noise, and low-drop segments during slow walking, producing a complete gait sequence.

For step-length estimation, the positive single-step acceleration area is compressed with a one-tenth power and fitted with a single-parameter power model. Least squares yields C=0.5124, reducing model complexity while supporting real-time per-step estimation.

The evaluation combines the self-collected SELF dataset with the public HAR dataset, covering 157 walking sequences across different participants, constant and variable speeds, indoor and outdoor surfaces, and sensor positions including the chest, forearm, and head.

The method achieves 99.73% overall mean step-counting accuracy with a standard deviation of 0.50%. On the step-length test set, it reaches a per-step MAE of 3.53 cm and a MAPE of 4.98%, with 93.17% of samples having errors no greater than 8 cm.

Category: 2026 National College Students Surveying and Mapping Innovation and Entrepreneurship Intelligence Competition, Scientific Paper Track.

In the project experiments, the single-feature, single-parameter model balanced step-counting accuracy, step-length performance, and computational cost. It provides gait-event and per-step displacement inputs for smartphone PDR; future work will extend validation to full trajectories through heading estimation and gyroscope integration.