<p>Human mobility analysis is vital for urban planning, public safety, and resource optimization in modern smart cities. Traditionally, mobility monitoring has relied on GPS tracking, CCTV, or mobile applications. Although these approaches are effective, they raise substantial privacy concerns. As public awareness of data privacy grows, demand for privacy-preserving monitoring approaches is increasing. One promising approach is footstep-based monitoring using seismic sensors, which detect ground vibrations induced by human movement without identifying personal information. This study examines the feasibility of a seismometer-based pedestrian monitoring system in a high-noise urban environment. The research utilized seismic data collected over seven distinct sessions at the University of Tokyo, representing varied ambient conditions, with a camera providing ground-truth validation for all measurements. Our two-stage methodology first uses spectral feature extraction, optimized with a 5-second analysis window and a 95% overlap ratio, to reliably detect human-induced vibrations. Then, a Convolutional Neural Network (CNN) incorporating spectral subtraction for noise reduction and trained with augmented spectrogram data was developed to classify pedestrian group size. The performance was evaluated by using the F1 score for detection and 5-fold cross-validation for classification. The optimized detection method yielded a robust F1-score exceeding 0.92, indicating high reliability. The regression-based CNN classifier achieved over 80% test accuracy and maintained over 72% classification accuracy on external continuous data. These findings demonstrate that the proposed system can successfully differentiate between pedestrian group sizes based on ground vibration signals, even in the presence of significant ambient noise. This study confirms the viability of seismometer-based monitoring as an effective, privacy-preserving tool for urban analytics, security surveillance, and unobtrusive monitoring.</p> Graphical Abstract <p>Based on the graphical snapshot, this study was conducted to determine the feasibility of using a seismometer for pedestrian monitoring as a privacy-preserving alternative to conventional methods like cameras. The research proposes a two-stage methodology to analyze urban mobility. The first stage focuses on pedestrian detection using spectral feature extraction from raw seismic data. This method was optimized using a 5-second analysis window and 95% overlap, achieving an F1-score of over 0.92 in noisy urban settings. The second stage involves counting and classifying pedestrians. A Convolutional Neural Network (CNN) was trained on spectrograms of the detected seismic signals. The model's performance was enhanced through techniques like spectral subtraction for noise removal and data augmentation. The CNN classifier successfully differentiates between pedestrian group sizes (solo, couple, and group). The results demonstrate the system's robust performance, with the CNN model achieving over 90% validation accuracy and maintaining over 80% classification accuracy on continuous data. The study concludes that seismometer-based monitoring is an effective, privacy-preserving tool for urban analytics and surveillance, offering a practical solution for urban planning and public safety.</p> <p></p>