Using LiDAR and IMU Fusion for Accurate Navigation in Mobile Robots
Mobile robots operating in indoor environments face the challenge of maintaining accurate pose estimation while navigating through corridors, rooms, and around obstacles. Standalone sensors often exhibit limitations: LiDAR provides precise range measurements but can suffer from featureless areas or dynamic occlusions, while inertial measurement units (IMUs) offer high-rate acceleration and angular velocity data that drifts over time. Combining these two sensing modalities through sensor fusion can compensate for individual weaknesses and produce more reliable state estimates.
The core idea behind LiDAR-IMU fusion is to integrate complementary data streams into a common estimation framework. LiDAR point clouds deliver absolute spatial references relative to the environment, which can correct the cumulative errors inherent in IMU dead-reckoning. Conversely, the IMU provides continuous motion information between LiDAR scans, enabling smoother trajectory reconstruction and better handling of rapid movements. This article examines two widely used estimation methods — the Kalman filter and the particle filter — and discusses how each approach handles fusion for indoor mapping applications.
Both filters have been studied extensively in robotics and are available in many open-source platforms. The choice between them depends on factors such as computational resources, environmental characteristics, and required accuracy. The following sections describe the principles behind each technique and outline considerations for practical implementation in an indoor mapping context.
Sensor Modalities and Their Characteristics
LiDAR sensors emit laser pulses and measure the time-of-flight to return, producing a point cloud that represents the geometry of surrounding surfaces. In indoor environments, LiDAR can capture walls, furniture, and other static features with high angular resolution and range accuracy. However, LiDAR data can become sparse in long, uniform corridors or when the robot moves through glass or mirror surfaces. Additionally, the update rate of most LiDARs (typically 10–20 Hz) may be insufficient for fast maneuvers.
IMUs contain accelerometers and gyroscopes that measure linear acceleration and angular velocity. These measurements are integrated over time to estimate velocity and orientation, but the process amplifies bias and noise, leading to unbounded drift. An IMU alone cannot provide absolute position information; it only tracks relative changes. In indoor settings, magnetic field disturbances may also affect orientation estimates from magnetometers, which are sometimes included in IMUs.
When fused, LiDAR and IMU data complement each other. The IMU fills in motion during intervals between LiDAR scans, allowing the filter to maintain a continuous state estimate. Meanwhile, each new LiDAR scan resets drift by aligning observed features with the map. This synergy is especially valuable for indoor mapping, where frequent loop closures are rare and the robot may revisit spaces after long periods.
Principles of Sensor Fusion for Pose Estimation
Sensor fusion methods combine measurements from multiple sources to produce a state estimate that is more accurate than any single sensor alone. In the context of LiDAR and IMU, the state typically includes the robot’s position, orientation (pose), linear velocity, and angular velocity, along with IMU bias terms. All sensors must be referenced to a common coordinate frame, and their timestamps must be synchronized to within a few milliseconds to avoid misalignment.
A typical fusion pipeline begins with preprocessing: LiDAR point clouds are often downsampled or filtered to reduce computational load, while IMU readings are calibrated and corrected for bias. The filter then predicts the state forward using the IMU data as a motion model, and updates the state when a LiDAR scan arrives by matching it against a local or global map. The matching step — often performed through scan-to-map registration — provides a measurement of the robot’s pose relative to the environment.
Both Kalman filters and particle filters rely on this prediction-update structure, but they differ in how they represent probability distributions over the state space. The Kalman filter models uncertainties as Gaussian distributions, while the particle filter uses a set of weighted samples (particles) to approximate arbitrary distributions. This distinction leads to different strengths and weaknesses in indoor mapping scenarios.
Kalman Filter Approach
The extended Kalman filter (EKF) is a common variant used for nonlinear systems such as robot navigation. It linearizes the motion and measurement models around the current estimate using Taylor expansion. In LiDAR-IMU fusion, the prediction step integrates IMU measurements to propagate the state and its covariance, while the update step corrects the state using the residual between predicted and observed LiDAR features.
One advantage of the EKF is its computational efficiency. Because the covariance matrix is propagated analytically, the update requires only matrix operations that scale quadratically with the state dimension. For typical mobile robot states (15–20 variables), an EKF can run at several hundred hertz on embedded hardware, leaving resources for other tasks such as mapping or planning. This makes it suitable for real-time applications where latency must be minimized.
However, the EKF’s Gaussian assumption can be limiting in indoor environments. When the robot’s pose distribution becomes multimodal — for example, during ambiguous alignment in symmetric hallways or after long periods of drift — the EKF tends to converge to one mode, potentially losing the true position. Moreover, linearization errors can accumulate, especially during rapid rotations or when the motion model is highly nonlinear. These limitations motivate the use of alternative filtering methods in certain scenarios.
Particle Filter Approach
Particle filters, also known as Monte Carlo localization (MCL) in robotics, represent the posterior distribution by a set of weighted particles. Each particle is a hypothesis of the full state, and the filter resamples particles according to their weights after each measurement update. For LiDAR-IMU fusion, the prediction step propagates each particle using IMU data (possibly with added noise to account for uncertainty), and the update step assigns weights based on how well the simulated LiDAR scan matches the actual scan.
This approach naturally handles non-Gaussian and multimodal distributions. In an indoor environment, if the robot is uncertain about its location — for instance, after being kidnapped or after traversing a long, featureless corridor — the particle filter can maintain multiple possible poses simultaneously. As more LiDAR scans arrive, particles with high consistency survive, and the distribution eventually collapses to the correct mode. This robustness is particularly valuable for global localization or recovery from tracking failures.
On the downside, particle filters are computationally more expensive than EKFs. The number of particles required increases exponentially with state dimensionality. For a full 3D pose (6 degrees of freedom) plus velocity and biases, tens of thousands of particles may be needed to maintain adequate coverage. Adaptive resampling strategies and efficient scan-matching algorithms can mitigate this cost, but in resource-constrained systems the Kalman filter often remains the preferred choice.
Comparative Analysis for Indoor Mapping
When comparing the two filters for indoor mapping tasks, trade-offs between accuracy, reliability, and computational load become evident. In structured environments with many distinctive features — such as rooms with furniture, doorways, and corners — the EKF typically delivers sufficient accuracy while using less processing time. The Gaussian assumption holds reasonably well because the distribution remains unimodal and the linearization error is small. Consequently, many commercial and research platforms employ EKF-based LiDAR-IMU fusion for real-time mapping.
In contrast, particle filters excel in scenarios where the robot faces ambiguity. For example, in a long, symmetric corridor with repeated textures, the EKF may drift laterally without any correction, whereas a particle filter can keep multiple hypotheses until a distinctive feature resolves the ambiguity. This makes particle filters a strong choice for multi-story buildings, open-plan offices, or environments with moving obstacles that cause temporary sensor occlusion. However, the increased computational demand may require more powerful processors or higher-level optimization.
Another consideration is map representation. Many indoor mapping systems, such as those using occupancy grids or signed distance fields, benefit from a continuous pose estimate provided by a filter. Both methods can be coupled with a mapping backend (e.g., a separate SLAM module) that updates the global map as new poses are generated. In practice, some systems use a hybrid approach: an EKF for local tracking and a particle filter for global localization or re-initialization.
Implementation Considerations
Deploying a LiDAR-IMU fusion system requires careful attention to sensor calibration and synchronization. The extrinsic calibration between the LiDAR and IMU — determining their relative translation and rotation — must be performed accurately, often using a calibration target or a batch optimization procedure. Without correct extrinsics, the fused estimates will contain systematic errors. RoboCore’s development kits include tools for such calibration, which can simplify the initial setup.
Timing alignment is equally critical. IMU data should be timestamped at capture, and LiDAR scans should be interpolated to match the IMU time base. Many modern LiDARs provide hardware-synchronized output with microsecond precision. Software solutions, such as using a dedicated real-time clock or timestamping over a bus, can also achieve acceptable synchronization for indoor applications. Additionally, the filter update rate must be balanced — updating on every LiDAR scan is typical, but partial updates (e.g., using only a subset of points) can reduce computational load without sacrificing accuracy.
Finally, the choice between Kalman and particle filters often depends on the specific mapping objective. For high-speed mapping in known environments, an EKF combined with efficient scan-matching (such as ICP or NDT) is often sufficient. For robust operation in unknown or changing indoor spaces, a particle filter provides greater resilience to failures. Developers should evaluate their robot’s motion profile, environment, and available computing power before selecting one over the other.
