This paper aims to improve the robustness and interpretability of driving primitive segmentation in noisy naturalistic car following data. To handle strong sensor noise and brief abrupt maneuvers, we design a Sticky-HDP-HSMM that imposes stickiness on the duration layer, so that one to two second noise-induced micro segments are suppressed while state dwell times remain explicitly modeled. The model uses Gaussian emissions and a time indexed HSMM forward recursion. It is applied to the SPMD naturalistic driving dataset sampled at 10 Hz, where a unified preprocessing pipeline selects valid car following episodes and applies stronger Savitzky–Golay smoothing to longitudinal acceleration, yielding 1 227 car following segments from 44 drivers.We conduct a controlled comparison with HDP-HMM and Sticky-HDP-HMM under identical data, preprocessing, hyperparameter selection and Gibbs sampling with a maximum of 200 iterations. The proposed method raises the log likelihood from about −4200 to about −180 within the first 10 iterations and attains the highest steady state log likelihood, exceeding the baselines by roughly 540 and 770. Qualitative and quantitative analyses show markedly fewer segments shorter than 2 s, a dominant segment length mode between 5 and 10 s, and boundaries that align more closely with extrema in acceleration, zero crossings in relative speed and shifts in headway. Using kernel density valley thresholds for longitudinal acceleration, relative speed and headway, we discretize the variables into behavior states that provide weak semantic supervision for downstream driving style analysis. The results demonstrate a favorable trade off between noise suppression and behavioral fidelity and yield a reproducible and interpretable pipeline for driving primitive extraction.