Physically Interpretable Spectral Features for Urban Mobility Classification Using Distributed Acoustic Sensing

A novel processing and classification framework for the automatic detection and classification of pedestrians, bicycles, and cars in urban environments is proposed and experimentally demonstrated. The proposed processing pipeline combines high-sensitivity distributed acoustic sensing, automatic trajectory extraction, physically interpretable spectral feature extraction, and supervised machine-learning-based pattern recognition. The technique leverages pre-existing telecommunication fibers installed beneath or close to sidewalks and roadways. Coherent optical time-domain reflectometry enhanced with optical pulse compression provides the sensitivity and spatial resolution required to detect mobility-induced vibrations using these non-dedicated fibers, thereby avoiding sensing-cable deployment or pavement retrofitting. A detailed analysis of the signals extracted along the detected trajectories is performed in the spatial-frequency and equivalent-frequency domains. Controlled walking, running, and cycling measurements, together with passenger-vehicle observations under normal traffic conditions, reveal reproducible spectral characteristics associated with the underlying pavement-loading mechanisms. In particular, the analysis identifies four complementary properties: spatial loading periodicity, periodic loading peak prominence, initial spectral decay, and broadband spectral-energy distribution. Seven physically interpretable features derived from these properties are subsequently supplied to a support vector machine classifier. The processing and classification framework is evaluated using measurements acquired from an existing optical fiber in a real urban environment containing simultaneous sidewalk and road activity. When classification is restricted to manually verified bicycle, car, and pedestrian trajectories, a balanced accuracy of 98.3% is achieved. These results demonstrate the potential of physically interpretable spectral analysis of DAS measurements for continuous, spatially distributed, and privacy-preserving urban mobility monitoring.