Centralized Fusion

All raw sensor data (e.g., pixel streams, lidar point clouds, IMU readings) is transmitted to a central fusion node. This node houses the complete state estimation algorithm (e.g., a Kalman filter, particle filter, or factor graph optimizer). It performs the critical tasks of data association (matching measurements to tracks), filter prediction, and filter update in one location. This centralization simplifies algorithm design but creates a computational bottleneck and single point of failure.

Because the central node has access to all raw, unprocessed data simultaneously, it can theoretically achieve Bayesian optimal estimation. It can consider the joint probability distribution of all measurements given the state, avoiding the approximations and information loss inherent in architectures that pre-process data locally. This makes it the gold standard for performance comparison, though this optimality assumes perfect models, known noise statistics, and no communication delays—conditions rarely met in real-world systems.

This architecture requires transmitting vast amounts of high-dimensional raw data over a network. For example:

• A single high-resolution camera can generate over 100 MB/s.
• A 64-beam lidar produces millions of points per second.
• Raw radar data is also voluminous. This creates a significant network bandwidth bottleneck, limiting scalability and increasing system cost and power consumption, especially in mobile or embedded applications like autonomous vehicles.

The fusion node must be a high-performance compute unit capable of running complex estimation algorithms on all incoming data streams in real-time. This often necessitates powerful GPUs or AI accelerators. The load scales linearly with the number and data rate of sensors. This contrasts with decentralized fusion, where computational load is distributed across sensor nodes. The central node's failure means a total loss of perception capability.

Having all data arrive at one point simplifies two key engineering challenges:

• Temporal Synchronization: The central node can apply software timestamp alignment to correct for minor clock drifts across sensors, though hardware triggers are still preferred for precision.
• Extrinsic Calibration: The process of determining the rigid transformation between sensor frames is centralized. The calibration parameters are stored and applied at the fusion node, ensuring consistency across all data processing paths.

Centralized fusion is often used in systems where communication bandwidth is plentiful and central compute is powerful.

• Research Prototypes & Simulation: Used as a baseline in academic papers and high-fidelity simulators (e.g., CARLA, NVIDIA DRIVE Sim) to establish performance ceilings.
• Industrial Inspection Stations: Fixed systems where sensors are co-located and connected via high-speed wired networks.
• Early Autonomous Vehicle Stacks: Many first-generation prototypes used centralized architectures with a trunk full of compute, before evolving to more distributed, zonal designs for production.

In self-driving cars, a centralized fusion computer (often called the perception engine) ingests raw streams from lidar, radar, cameras, and ultrasonic sensors. This architecture is critical because:

• It enables optimal global data association, ensuring a camera detection of a pedestrian is correctly fused with a lidar point cloud and radar velocity measurement of the same object.
• It allows for complex, joint probabilistic models (like a centralized Kalman filter or factor graph) that consider all sensor uncertainties and correlations simultaneously.
• It provides a single, authoritative world model for the planning stack, avoiding inconsistencies that can arise in decentralized systems.

Command and control centers for air defense or wide-area surveillance utilize centralized fusion to maintain a single integrated air picture (SIAP) or common operational picture (COP).

• Raw data from distributed radars, electro-optical/infrared (EO/IR) sensors, and electronic support measures (ESM) are transmitted to a central fusion node.
• This node performs multi-sensor, multi-target tracking using algorithms like the Joint Probabilistic Data Association (JPDA) filter or Multiple Hypothesis Tracker (MHT), which are computationally intensive and require a global view of all measurements to resolve ambiguities in cluttered environments.
• It enables sensor resource management, where the central node can task specific sensors based on the global picture.

Predictive maintenance systems in factories or power plants use centralized fusion to monitor complex machinery. Vibration sensors, acoustic emission sensors, thermal cameras, and current/power meters stream data to a central analytics server.

• The server fuses these multi-modal time-series data to create a unified health signature for a turbine, pump, or bearing.
• A centralized model can learn subtle, cross-modal failure precursors—e.g., a specific vibration frequency appearing concurrently with a slight thermal anomaly—that would be invisible to individual sensor analyses.
• This supports fault detection and isolation (FDI) by correlating symptoms across the entire sensor suite.

Advanced surgical navigation systems and diagnostic imaging platforms fuse data centrally. During neurosurgery, for example, pre-operative MRI scans, real-time intraoperative ultrasound, and optical tracking of surgical instruments are fused into a single 3D visualization.

• A central workstation performs the non-rigid registration of these modalities, compensating for brain shift during surgery.
• This requires processing the raw or minimally processed image data centrally to achieve the sub-millimetric accuracy needed for safe navigation. The fusion provides a augmented reality view for the surgeon, overlaying critical structures from the MRI onto the live surgical field.

Modern aircraft integrated modular avionics (IMA) systems often employ a centralized fusion core. Data from the inertial navigation system (INS), GPS, air data computers, and radio altimeters are fed to a central flight control computer or navigation filter.

• This central unit runs a sophisticated estimator (like an extended Kalman filter) to compute the aircraft's precise attitude, position, and velocity (state estimation).
• The architecture ensures fault tolerance through centralized voting and comparison; if one sensor disagrees with the fused consensus, it can be isolated. The resulting best estimate of the state is then distributed to all downstream systems (displays, autopilot, etc.).

An alternative architecture where each sensor node performs local processing and state estimation before sharing results with other nodes to achieve a global consensus. Unlike centralized fusion, there is no single point of failure or bottleneck. This design is common in distributed systems like swarm robotics or ad-hoc sensor networks where communication bandwidth is limited or reliability is paramount. Key challenges include managing communication latency and ensuring data association consistency across nodes.

A recursive algorithm that is the mathematical foundation for many centralized fusion systems. It optimally estimates the state of a linear dynamic system from a series of noisy measurements. The filter operates in a two-step cycle:

• Prediction: Projects the current state forward using a process model.
• Update: Corrects the prediction with a new measurement. It maintains a covariance matrix representing estimation uncertainty. In centralized fusion, a single Kalman filter (or its nonlinear variants) often fuses all raw sensor data at the central node.

The overarching goal of sensor fusion: to infer the hidden variables (the state) of a dynamic system from noisy, incomplete sensor observations. The state typically includes quantities like position, velocity, and orientation. Centralized fusion is a specific architectural approach to solving the state estimation problem. Other related problems within this domain include:

• Smoothing: Estimating past states using future measurements.
• Prediction: Forecasting future states. The accuracy of state estimation directly determines the performance of autonomous systems.

A mathematical representation that defines how a sensor's measurements relate to the true state of the world. It is a critical component of any fusion architecture. A sensor model includes:

• Observation function: Maps the system state to the expected sensor reading.
• Noise characteristics: Typically modeled as Gaussian with a known covariance matrix.
• Field of view and detection probabilities. In centralized fusion, the central node uses precise sensor models for each incoming data stream to correctly interpret and combine the raw measurements.

A square matrix that quantifies the uncertainty and correlation of a state estimate. Each diagonal element represents the variance (uncertainty) of an individual state variable (e.g., x-position). Off-diagonal elements represent the covariance (correlation) between variables (e.g., how uncertainty in x-position relates to uncertainty in x-velocity). In centralized Kalman filtering, the covariance matrix is propagated and updated with each prediction and measurement step, providing a rigorous measure of the fused estimate's confidence.

The challenging sub-problem of determining which incoming sensor measurements correspond to which real-world objects or existing tracks. It is a prerequisite for effective fusion. In a centralized architecture, the central node must solve this for all sensors simultaneously. Algorithms for this include:

• Nearest Neighbor with gating using the Mahalanobis distance.
• Joint Probabilistic Data Association (JPDA).
• Multiple Hypothesis Tracking (MHT). Poor data association leads to track mixing or false tracks, severely degrading system performance.