知识点

21. What is drift in SLAM?

What is scale drift?

✅ Sample Interview Answer:

In SLAM, drift refers to the gradual accumulation of error in the estimated pose of the robot over time due to small inaccuracies in sensor measurements or motion estimation. Since SLAM is often based on integrating relative measurements (like odometry or visual feature tracking), even tiny errors can build up, causing the estimated position to diverge from the true position.

Scale drift, on the other hand, is a specific kind of drift that occurs mainly in monocular visual SLAM, where the system does not have direct access to real-world scale (since a single camera can't measure depth). Over time, the estimated scale may stretch or shrink inconsistently, making distances or map size inaccurate. This can result in misinterpretation of spatial relationships in the generated map.

To mitigate drift, SLAM systems often incorporate loop closure detection and global optimization.

23. What is relocalization in SLAM?

Relocalization is the process of recovering the correct pose of a robot or camera within a previously built map after tracking has been lost. This can happen due to motion blur, occlusion, or system restart.
🔹 Goal: Re-establish the current pose using existing map data.

How does relocalization differ from loop closure detection?

Aspect Relocalization Loop Closure Detection
When After tracking is lost or upon re-entering a known area During continuous tracking
Why To resume pose estimation To correct accumulated trajectory drift
Trigger Loss of tracking or restart Revisiting a previously seen location
Result System can continue localization Adds a constraint to pose graph for map optimization
Analogy “Where am I now?” “I’ve been here before—let’s refine my path.”

✅ Sample Interview Answer:

“Relocalization in SLAM is used when the system loses tracking and needs to determine its current pose in a previously mapped environment. It's different from loop closure, which happens during normal operation when the system recognizes a place it's been before to correct accumulated drift. Relocalization answers the question ‘Where am I now?’, while loop closure says ‘I’ve been here before—let’s adjust the map for better accuracy.’”

LiDAR SLAM

  1. Explain how the Iterative Closest Point (ICP) algorithm functions.

    • Which derivative work of the ICP algorithm do you prefer and why?

    🔄 1. How ICP Works

    1. Initialization
      Choose one cloud as the fixed target, and the other as the movable source. Optionally, start with an initial guess (pose).
    2. Closest-Point Matching
      For each point in the source, find the nearest neighbor in the target, typically using a K‑D tree for efficiency.
    3. Transformation Estimation
      Estimate the best rigid transform—rotation R and translation t—to minimize an error metric (e.g. sum of squared distances in point‑to‑point or point‑to‑plane fashion).
    4. Apply Transformation
      Update the source cloud by applying the computed R and t.
    5. Convergence Check or Iterate
      Evaluate whether the change in mean error is small enough; if not, repeat steps 2–5.

  2. Discuss the Point-to-Point and Point-to-Plane metrics in the context of the ICP algorithm.

  3. If an ICP algorithm fails to register two sets of point clouds, what could be the possible reasons?

🛠️ Practical Checklist to Diagnose and Fix Failures

Problem Remedy
Poor initial pose Use global registration or feature-based alignment
Partial overlap Crop to overlap, trim unmatched regions
Noise / outliers Apply filters (e.g., Statistical Outlier Removal)
Symmetric or ambiguous structures Add distinctive features or constraints
Sparse point clouds Densify or increase resolution sampling
Bad normals Recompute normals with better parameters

Final Insights

ICP is robust within its assumptions, but can fail when those assumptions aren’t met—like insufficient overlap, noise, symmetry, or poor initial alignment. For reliable results:

  1. Ensure a good initial guess (possibly via global registration).

  2. Preprocess for quality point clouds (noise/outlier removal, densification).

  3. Use robust variants (e.g., point-to-plane, weighted ICP).

  4. Provide feature-rich inputs—add constraints or landmarks.