The paper presents a framework for modeling and predicting short‑term human dynamic behavior to enable adaptive human‑machine control systems. It treats a human operator as a Markov device with multiple internal "mental" states, each associated with a distinct dynamic controller. The authors describe three modeling layers: (1) simple linear dynamic models estimated with Kalman filters; (2) multiple‑model (generalized likelihood) approaches that select among several Kalman filters based on measurement innovations; and (3) Hidden Markov Dynamic Models (HMDM) that incorporate explicit state transition probabilities to capture ordered sequences of internal states. The methodology is applied to a driving‑simulator experiment using a Nissan 240SX platform with eight subjects performing six command types (stop, turn left/right, lane change, pass, drive‑normal). Steering and acceleration data are used to train three‑state HMDMs for each action, and real‑time Viterbi decoding classifies driver intent within one second of command presentation, achieving 95 % mean recognition accuracy. An appendix details a kinematic‑dynamic car model used in the simulator, including equations for position, heading, velocity, energy balance, and skid detection. The authors conclude that early intent recognition via HMDMs can enable control systems that dynamically complement human actions and anticipate broader applicability to other human‑machine domains.
Topics
Human as a Markov Device: Internal State Inference for Adaptive Control
Conceptualization of the human operator as a multi‑state Markov process, where each internal state governs a distinct control behavior and transitions are probabilistic.
Simple Linear Dynamic Models and Kalman Filtering for State Estimation
Formulation of single‑process state evolution and observation equations, and the use of optimal linear Kalman filters to estimate and predict the human’s state vector.
Multiple‑Model (Generalized Likelihood) Approach for Prototypical Behaviors
Extension to a bank of Kalman filters, each representing a prototypical driver style; selection based on measurement‑innovation likelihoods to identify the currently active model.
Hidden Markov Dynamic Models (HMDM) for Sequential Intent Modeling
Integration of explicit Markov transition probabilities among dynamic models, continuous‑density HMM formulation, and Viterbi decoding to infer ordered internal states underlying actions.
Driving‑Simulator Experiment: Real‑Time Intent Recognition
Design, data collection, and evaluation of a Nissan 240SX simulator study with eight subjects performing six command types; training of three‑state HMDMs and reporting of 95 % recognition accuracy at 1 s post‑command.
Dynamic Car Model Used for Simulation
Kinematic and energetic equations governing vehicle position, heading, velocity, drag, braking, and skid detection, calibrated for a Nissan 240SX.
Implications for Adaptive Human‑Machine Systems
Discussion of how early intent detection enables control systems to reconfigure dynamically, with potential generalization beyond driving to other human‑machine interactions.