Process Model

A process model is formally defined as a state transition function. For a discrete-time system, it predicts the next state (x_k) from the previous state (x_{k-1}) and a control input (u_k), subject to process noise (w_k) which accounts for unmodeled dynamics and disturbances.

• General Form: (x_k = f(x_{k-1}, u_k, w_k))
• Linear Example (Kinematic): For a constant velocity model, the state (x = [p, v]^T) (position, velocity) evolves as: (p_k = p_{k-1} + v_{k-1} \Delta t), (v_k = v_{k-1} + w_k).
• The function (f(\cdot)) encapsulates the system's physics, such as Newtonian mechanics for robots or aerodynamics for drones.

In Bayesian filtering, the process model is represented as a probability distribution: the state transition probability (p(x_k | x_{k-1}, u_k)). This distribution quantifies the likelihood of transitioning to a new state, given the previous state and control.

• This probabilistic view explicitly handles uncertainty. Instead of a single predicted state, it yields a predicted belief distribution.
• For the Kalman Filter, this is a Gaussian: (p(x_k | x_{k-1}) = \mathcal{N}(x_k; F x_{k-1}, Q)), where (F) is the state transition matrix and (Q) is the process noise covariance.
• For a Particle Filter, this distribution is used to propagate (move) each particle forward, diffusing them according to the model's uncertainty.

The primary function of the process model is to drive the prediction step (also called the time update) in recursive Bayesian filters like the Kalman Filter, Particle Filter, and their variants.

• Sequential Process: 1. Start with a prior belief (p(x_{k-1})). 2. Apply the process model (p(x_k | x_{k-1}, u_k)) to compute the predicted prior (p(x_k)) for the new timestep.
• This step propagates uncertainty forward. Even with a perfect previous estimate, process noise injects new uncertainty, causing the belief to become less certain (covariance increases).
• The prediction is then corrected by the update step using fresh sensor measurements.

A critical characteristic is whether the model is linear or nonlinear, which dictates the choice of estimation algorithm.

• Linear Models: Expressed as (x_k = F x_{k-1} + B u_k + w_k). These are optimally handled by the standard Kalman Filter.
• Nonlinear Models: Involve functions like trigonometry (e.g., for orientation), aerodynamic drag, or complex kinematics. Examples include:
  • Robot turning: (\theta_k = \theta_{k-1} + (\omega \Delta t))
  • Drone dynamics with Euler angles.
• Nonlinear models require approximations: the Extended Kalman Filter (EKF) uses Jacobians to linearize, while the Unscented Kalman Filter (UKF) uses deterministic sampling.

The process noise covariance matrix (Q) is a fundamental parameter of the model that quantifies the confidence in the model's own predictions. It is the covariance of the noise term (w_k).

• What it represents: Uncertainty from unmodeled forces, approximations in the dynamics, discretization errors, and unknown external disturbances (e.g., wind gusts).
• Tuning (Q) is crucial for filter performance. A (Q) set too low makes the filter overconfident in its model, causing it to lag behind true state changes. A (Q) set too high makes the filter rely more on noisy measurements, increasing jitter.
• In practice, (Q) is often tuned empirically or derived from known physical bounds on acceleration or jerk.

Specific process models are standardized for common tracking and navigation scenarios:

• Constant Velocity (CV): Assumes no acceleration. State: ([x, y, \dot{x}, \dot{y}]). Simple but effective for smooth motion.
• Constant Acceleration (CA): Includes acceleration as part of the state: ([x, y, \dot{x}, \dot{y}, \ddot{x}, \ddot{y}]). More responsive but noisier.
• Coordinated Turn (CT): Models turning motion with a (nearly) constant turn rate (\omega). Critical for tracking aircraft or maneuvering vehicles.
• Unicycle/Differential Drive: For wheeled robots, linking wheel velocities to linear and angular robot velocity.
• IMU Kinematic Model: Uses accelerometer and gyroscope measurements as control inputs (u_k) to propagate a pose (position, velocity, orientation).

Process models predict a vehicle's future position, velocity, and orientation based on control inputs (steering, throttle) and vehicle dynamics. This prediction is fused with real-time GPS, IMU, and wheel odometry data in a Kalman filter to generate a high-frequency, accurate pose estimate, even when GPS signals are temporarily lost in urban canyons or tunnels.

Quadcopters and fixed-wing drones use process models (often kinematic or dynamic models) to predict attitude and position. This prediction is combined with data from gyroscopes, accelerometers, magnetometers, and barometers to stabilize flight. Advanced systems integrate Visual-Inertial Odometry (VIO), where the process model helps fuse camera images with IMU data for precise, GPS-denied navigation indoors or in complex environments.

In industrial automation, a robotic arm's dynamic process model predicts the motion of its joints and end-effector given motor torques. This model is used within a control loop to:

• Predict the arm's state milliseconds ahead.
• Fuse predictions with encoder feedback for precise, smooth trajectory execution.
• Compensate for dynamic effects like inertia and friction, enabling high-speed, high-precision tasks like welding or assembly.

The high-fidelity orbital mechanics models used for satellites are complex process models accounting for gravity, atmospheric drag, and solar radiation pressure. Ground stations receive noisy range and Doppler measurements. A filter (like an Extended Kalman Filter) uses the orbital process model to predict the satellite's state, then corrects it with measurements to maintain a precise, continuously updated estimate of its position and velocity for station-keeping and mission operations.

In perception systems for robots and cars, process models track the motion of other dynamic agents. Simple constant velocity or constant acceleration models predict where a detected pedestrian or vehicle will be in the next sensor frame (e.g., from a lidar or camera). These predictions are crucial for the data association step in Multi-Object Tracking (MOT), linking detections across time to form stable tracks and estimate trajectories for path planning and collision avoidance.

An INS is a classic application of dead reckoning via a process model. It double-integrates accelerometer readings (using a model of gravity) to estimate position and integrates gyroscope readings to estimate orientation. Because tiny sensor errors accumulate rapidly, the INS process model's prediction is continuously corrected by fusing with absolute reference sensors like GPS in a tightly coupled or loosely coupled sensor fusion architecture. This provides smooth, high-bandwidth navigation.