CamVox: A Low-cost and Accurate Lidar-assisted Visual SLAM System

I. INTRODUCTION

• CamVox combines camera input with Livox lidar in an ORB-SLAM2-based visual SLAM system.The paper presents it as a low-cost lidar-assisted system.Source: §I · p. 1
• Its preprocessing synchronizes lidar and camera data, uses IMU data to correct scan distortions, and transforms lidar points into camera frames.The system also uses Livox’s non-repeating scan pattern for automatic camera–lidar calibration in uncontrolled scenes.Source: §I · p. 1
• The authors report that CamVox supports large-scale mapping with higher efficiency and robust operation in strong sunlight outdoors.They also describe its lidar point-cloud accuracy and range as superior to other depth cameras.Source: §I · p. 1
• The paper evaluates CamVox against mainstream frameworks and reports very accurate trajectory results.It also open-sources the hardware, code, and dataset as an out-of-the-box solution.Source: §I · p. 1
• The example map’s long side is about 200 meters, and its figure includes a robot trajectory, dense RGB-D map, and magnified point-cloud detail.The caption identifies the magnified view as demonstrating point-cloud quality.Source: §I · p. 1, Figure 1 · p. 1

II. RELATED WORK

• Monocular ORB-SLAM cannot recover real-world scale accurately, while ORB-SLAM2 adds stereo and RGB-D depth support.Stereo long-depth accuracy depends on a long baseline and is susceptible to calibration changes; RGB-D cameras are sunlight-sensitive and typically range less than 10 meters.Source: §II · pp. 1–2
• Lidar provides direct spatial measurements and can be robust in dynamic environments, but its lower angular resolution can hinder performance in tunnels or hallways, and LOAM lacks loop closure.LeGO-LOAM adds loop closure and feature-extraction changes; other camera–lidar fusion approaches include loosely coupled and tightly coupled systems.Source: §II · p. 2
• Livox’s prism-based, non-repeating scans can build high-definition point clouds as they accumulate; Horizon reaches 64 lines at 100 ms and detects up to 260 m under strong sunlight.The paper notes that the scan pattern may help camera–lidar calibration where traditional multiline lidar lacks precision between lines.Source: §II · p. 2
• Camera–lidar calibration remains challenging, with prior approaches using targets or manual effort and no open-source algorithm identified for uncontrolled scenes.The paper points to Livox’s non-repeating scan pattern as a possible solution.Source: §II · pp. 2–3

III. CAMVOX FRAMEWORK

• CamVox builds on ORB-SLAM2’s RGB-D model, with separate RGB-D input preprocessing and automatic calibration for uncontrolled scenes.Its platform includes a camera, Livox Horizon, and IMU; GPS/RTK is used for ground-truth estimation.Source: §III · p. 3, Figure 3 · p. 3
• Lidar-assisted visual keyframes generate local maps, while ORB-SLAM2’s backend uses pose-graph optimization, bundle adjustment, and loop closure.The paper describes this lightweight optimization as contributing to robustness.Source: §III · p. 3
• The platform figure shows the robot-mounted CamVox hardware, an additional RGB-D camera for comparison, and example RGB and lidar-derived depth images.The close-up identifies the camera, Livox Horizon, and IMU.Source: Figure 3 · p. 3

A. Hardware and software

• The camera, Livox Horizon lidar, and IMU are hard-synchronized by a 10 Hz trigger; the IMU runs at 200 Hz, and the lidar outputs continuously timestamped points.The system also records GPS-RTK data for ground-truth comparison and mounts a Realsense D435 for comparison.Source: A. Hardware and software · p. 3
• A preprocessing thread combines synchronized camera and IMU-corrected lidar data into a unified RGBD frame, alongside ORB-SLAM2’s major threads and a camera–lidar calibration thread.Calibration can start automatically when the robot is detected to be stationary or through human interaction.Source: A. Hardware and software · p. 3, Figure 4 · p. 3
• IMU-measured motion transforms lidar points from their individual scan times into the lidar coordinate frame at the camera-trigger time, correcting scan distortion.Fig. 5 contrasts a raw moving scan with the IMU-corrected point cloud.Source: A. Hardware and software · p. 4, Figure 5 · p. 4
• CamVox reliably detects depth-associated camera features beyond 100 meters, while the Realsense RGBD camera does not detect beyond 10 meters and suffers from sunlight noise.The system uses close keypoints with associated depth below 130 meters.Source: A. Hardware and software · p. 4, Figure 6 · p. 4

C. Calibration

• Calibration can run in uncontrolled scenes after the robot is detected to be stationary, accumulating lidar scans for a few seconds to form a dense depth image for correspondence with the camera image.The method addresses the need to recalibrate after possible misalignment and avoids motion blur during accumulation.Source: C. Calibration · p. 4
• The method projects lidar depth and reflectivity onto the imaging plane using initial extrinsics, then compares the resulting contours with camera-image contours and optimizes extrinsics with an improved ICP cost.The optimized cost is evaluated against the previous value to decide whether to update the extrinsic parameters.Source: C. Calibration · p. 4, Figure 7 · p. 4
• After histogram equalization, Canny edges from lidar depth and reflectivity are combined, short or cluttered interior edges are filtered, and shared camera–lidar edges are matched by nearest-edge distance.Edges shorter than 200 pixels are removed.Source: C. Calibration · p. 5
• The ICP cost adds a degree-of-mismatch term to address wrong matches in which very few points contribute, and coordinate descent iteratively optimizes roll, pitch, and yaw.The authors report that a value of 10 for the weighting factor is a good starting default.Source: C. Calibration · p. 5

IV. RESULTS

• The evaluation presents CamVox trajectory results in comparison with mainstream SLAM frameworks.The passage does not report the comparison outcome.Source: §IV · p. 5
• The evaluation examines the results of automatic calibration.Source: §IV · p. 5
• The evaluation studies the effect of the depth threshold used for close keypoints and includes a time analysis.Source: §IV · p. 5

A. Automatic calibration results

• The automatic calibration cost was 6.11, close to the best manual result of 5.88 and below the uncalibrated value of 7.95.Fig. 8 compares the uncalibrated, automatically calibrated, and best manual overlays.Source: A. Automatic calibration results · p. 5, Figure 8 · p. 5
• Calibration was triggered after an initial overlay misalignment of more than 2 degrees.The uncalibrated overlay is shown in Fig. 8(a), followed by automatic calibration in Fig. 8(b).Source: A. Automatic calibration results · p. 5, Figure 8 · p. 5
• The image-like calibration scheme worked robustly on most uncontrolled scenes, with evaluated scene costs around or below 6.Fig. 8 includes outdoor roads, outdoor artificial structures, and indoor underexposed structures.Source: A. Automatic calibration results · p. 5, Figure 8 · p. 5

B. Evaluation of depth threshold for keypoints

• Raising the close-keypoint depth threshold from 20m to 130m significantly increased mapping scale and mapped features.The lidar detects up to 260 meters, providing many fused-frame keypoints that can be treated as close.Source: B. Evaluation of depth threshold for… · p. 5
• Matching-point counts increased as more frames were captured during the first 100 frames at 10 FPS.The evaluation tracked matching points as a function of time after CamVox started.Source: B. Evaluation of depth threshold for… · p. 5
• A larger threshold tracked more features initially, and CamVox sets the default close-keypoint threshold to 130 meters.The authors say the additional initial matches help start robust localization and mapping.Source: B. Evaluation of depth threshold for… · p. 5

C. Comparison of trajectory

• The trajectory comparison used the SUSTech dataset and evo tools.The evaluation compared CamVox with two mainstream SLAM frameworks and ground truth.Source: C. Comparison of trajectory · p. 5
• Fig. 9 compares reconstructed maps across close-keypoint thresholds and plots matching-point counts by frame sequence and threshold.The threshold-based matching-count plot evaluates the beginning of SLAM at the fifth frame.Source: Figure 9 · p. 6
• CamVox was very close to ground truth and significantly outperformed Livox Horizon LOAM and VINS-mono.Fig. 10 shows the three trajectories alongside ground truth.Source: C. Comparison of trajectory · pp. 5–6, Figure 10 · p. 6

D. Timing results

• CamVox performed in real time on the onboard Manifold 2C with a 4-core Intel Core i7-8550U processor.The timing analysis is identified as Table II.Source: D. Timing results · p. 6, Table II · p. 6
• Automatic calibration took about 58 seconds to finish.Table II presents the framework timing analysis.Source: D. Timing results · p. 6, Table II · p. 6
• The calibration thread ran occasionally while the robot was stationary, and parameter updates could be deferred.The authors state this timing would not affect real-time performance.Source: D. Timing results · p. 6

• CamVox combines camera angular resolution with lidar depth and range in a low-cost visual SLAM framework.Source: §V · p. 6
• The Livox lidar working principle supports an automatic calibration algorithm designed for uncontrolled scenes.Source: §V · p. 6
• The framework was evaluated for calibration accuracy, close-keypoint depth-threshold classification, and trajectory comparison.Source: §V · p. 6
• CamVox runs in real time on an onboard computer.Source: §V · p. 6
• The authors propose CamVox as a potential low-cost, out-of-the-box solution for robotics and sensor research.Source: §V · p. 6