An accurate smartphone-based step-counting algorithm based on human walking acceleration variation patterns

Step count is a fundamental input for distance propagation in indoor pedestrian dead reckoning (PDR) and is widely used in mobile health monitoring and activity assessment. However, inertial sensor noise makes conventional smartphone-based step-counting algorithms, which often rely on fixed amplitude thresholds and temporal constraints, vulnerable to inter-subject gait differences, walking-speed variations, and disturbances during stationary-to-walking and walking-to-stationary transitions. These factors can cause missed steps and false counts in non-stationary scenarios such as speed-varying walking and stop-and-go transitions. To address these issues, this paper proposes an accurate smartphone-based step-counting algorithm based on walking-acceleration variation patterns. The method uses raw tri-axial acceleration signals and constructs a framework centered on Drop peaks, speed-change discrimination, and step-frequency-constrained denoising, rather than filtering or smoothing as the main preprocessing step. First, according to the upper frequency limit of fast human walking, each peak is paired with the minimum value in its corresponding region, and their difference is defined as the Drop value. Drop peaks are extracted as candidate step events, reducing the risk of mismatch from fixed amplitude thresholds. Second, acceleration, deceleration, and constant-speed walking are identified using adjacent Drop-peak-based amplitude changes and inter-step interval variations. Candidate events are screened and de-duplicated with an adaptive time-varying window to improve stability under speed changes. Finally, sequence-level step-frequency plausibility constraints suppress false triggers during start-stop transitions. Experiments on three datasets covering multiple gait patterns and indoor/outdoor surface conditions show an average step-counting accuracy of 99.75%. Comparative results demonstrate lower counting error and stronger robustness than representative baseline methods.

Designed a Drop-peak-based gait event extraction algorithm that dynamically characterizes acceleration variation through peak-valley differences, reducing missed and false counts caused by fixed thresholds.

Proposed a speed-state adaptive discrimination strategy that constructs a time-varying window combining magnitude and step-interval changes to filter and deduplicate candidate gaits.

Built a step-frequency-constrained sequence denoising method that suppresses false steps during start-stop phases and supplies the detected steps as distance-propagation input for indoor PDR.

The average step-counting accuracy across three datasets is 99.75%, with lower counting error than the representative baseline methods reported in the paper.

The method was evaluated under variable-speed walking, start-stop transitions, and different indoor and outdoor surface conditions.

Participated throughout the project from its initial stage, with responsibility for data analysis, algorithm design, robustness experiments, result visualization, manuscript writing, submission, and revision.