Sensor Fusion (TSRT14)
Course Content
The general problem is to estimate a state or parameter \(x\) from a multitude of sensor observations \(y\) distributed over time, space and modality (i.e., type of sensor). The course and book are structured as follows.
The static case (\(x\) is constant)
- General fusion theory
- Estimation theory in the linear case (\(y=Hx+e\)). Extensions to the non-linear case (\(y=H(x)+e\), or \(H(y,x,e)=0\)).
- Bayesian fusion perspective, the sensor fusion formula.
- Computational estimation issues: Centralized versus decentralized fusion (information propagation and double-counting, covariance intersection techniques). Batch computations versus sensor iterations.
- Detection theory \(T(x)>h\): likelihood ratio concepts, Neyman-Pearson, GLRT and other tests.
- Computational detection issues: Centralized and distributed detection. Censoring sensors.
- Diagnosis \(H(i): x=x(i)\): The vector model. Error approximations.
- Localization of a target based on one snap-shot from available
sensors
- Sensor networks: Range and range-difference measurements. Triangulation from bearings. Information limitations and censoring sensors.
- Measurement to target association, and extended target models (the fusion before detection principle).
The dynamic case (x is time-varying)
- Filter theory
- General non-linear filter theory for \(dx/dt=f(x,u,v)\), \(y=h(x,u,e)\): Numeric evaluation using the point mass filter. Two special cases (KF and HMM) and fundamental limitations (CRLB). Structured models and marginalization.
- Kalman filtering: Basic theory, implementation aspects, practical issues, information filter, smoothing).
- Approximative algorithms for non-linear models (Extended KF and HMM, UKF and sigma-point filters, KF banks).
- The particle filter: theory, implementation, proposal methods, sampling principles, smoothing, practical aspects.
- Time synchronization and coordinate system calibration in filtering.
- Dynamic state dimension problems
- Multi-target tracking: association, track handling
- Simultaneous localization and tracking (SLAM).
Practice
- Sensors, sensor models \(H(y,x,e)=0\) and sensor-near signal
processing
- Wheel speed sensors and odometry.
- IMU and dead-reckoning.
- GPS.
- Camera
- Radar, laser, and sonar
- Networked sensors: radio measurements, microphones, seismometers, magnetic field sensors
- Motion models \(dx/dt=f(x,u,v)\)
- Multi-purpose motion models
- Standard models for different platforms (wheeled vehicles, surface and underwater vessels, aircraft, …).
- Extended target models (track before detect, grid based methods for fusion)