# FAST-LIVO2 on Resource-Constrained Platforms: LiDAR-Inertial-Visual Odometry With Efficient Memory and Computation
Bingyang Zhou $^{ID}$, Chunran Zheng $^{ID}$, Member, IEEE, Ziming Wang $^{ID}$, Fangcheng Zhu $^{ID}$, Graduate Student Member, IEEE, Yixi Cai $^{ID}$, and Fu Zhang $^{ID}$, Senior Member, IEEE
Abstract—This paper presents a lightweight LiDAR-inertial-visual odometry system optimized for resource-constrained platforms. It integrates a degeneration-aware adaptive visual frame selector into error-state iterated Kalman filter (ESIKF) with sequential updates, improving computation efficiency markedly while maintaining a similar level of robustness. Additionally, a memory-efficient mapping structure combining a locally unified visual-LiDAR map and a long-term visual map achieves a good trade-off between performance and memory usage. Extensive experiments on x86 and ARM platforms demonstrate the system's robustness and efficiency. On the Hilti dataset, our system achieves a 33% reduction in per-frame runtime and 47% lower memory usage compared to FAST-LIVO2, with only a 3 cm increase in RMSE. Despite this slight accuracy trade-off, our system remains competitive, outperforming state-of-the-art (SOTA) LIO methods such as FAST-LIO2 and most existing LIVO systems. These results validate the system's capability for scalable deployment on resource-constrained edge computing platforms.
Index Terms—State estimation, sensor fusion, simultaneous localization, mapping.
### I. INTRODUCTION
DOMETRY [^1], [^2], [^3], [^4], [^5], [^6], [^7] has become an essential component of robotics applications, enabling autonomous systems to simultaneously map their environments and localize themselves. It plays a pivotal role in tasks such as trajectory planning [^8] and motion control [^9]. Recent advancements in odometry algorithms have substantially improved their accuracy, with many frameworks leveraging single-modality sensors like cameras or LiDAR. However, while visual and LiDAR odometry systems excel in specific domains, each faces inherent limitations that restrict their performance in diverse scenarios, such as LiDAR struggling in structureless environments and cameras failing in textureless scenes.
To address these challenges, multi-sensor fusion approaches, particularly LiDAR-inertial-visual odometry (LIVO), have excelled in tackling the limitations of single-sensor systems. These methods integrate complementary sensors' data to enhance localization robustness across various environments. However, most current LIVO systems [^3], [^7] rely on computationally intensive processes, necessitating high-performance CPUs to meet the real-time demands of sensor fusion, mapping, and optimization. The SOTA LIVO system, FAST-LIVO2 [^5], employs direct methods for state estimation leveraging all LiDAR and visual measurements within a sequential update ESIKF framework, and integrates a unified LiDAR-visual local map, achieving remarkable levels of computational efficiency and accuracy. Nevertheless, LIVO systems generally require intensive computational resources. Although some state-of-the-art LIVO systems, such as FAST-LIVO2, have demonstrated good efficiency, they still present computational demands that can be further optimized.
As robotics applications expand to lightweight and portable platforms (e.g., drones, autonomous vehicles, and IoT devices), deploying odometry on edge computing platforms becomes increasingly critical. Edge platforms, such as ARM architectures, offer advantages like energy efficiency and widespread adoption in embedded systems, making them attractive for scalable and cost-effective solutions. However, their constrained computational power, memory, and energy efficiency impose marked challenges to traditional odometry frameworks to ensure resource-efficient performance. Despite these challenges, edge computing brings unique benefits for real-time odometry, such as reduced latency and network dependency through localized data processing.
To enable more efficient odometry on edge computing platforms, we improve the state of the LIVO system, FAST-LIVO2, by minimizing its computational and memory overhead while maintaining its robustness and performances Fig. 1. The primary contributions of this work, when compared to FAST-LIVO2, are as follows:
1) Efficient visual update via LiDAR-degeneration-aware design: Rather than naturally using all available visual measurements, a LiDAR-degeneration-aware adaptive visual frame selector is integrated into the ESIKF framework of FAST-LIVO2, significantly reducing the computational
---
burden while maintaining a similar level of odometry accuracy.
2) Memory-efficient hybrid map structure: we propose a hybrid map structure that combines a compact local visual-LiDAR map with a long-term visual map. The latter stores sparse historical image observations to ensure odometry robustness while allowing the former to be maintained at a markedly smaller scale, reducing memory overhead. Additionally, the adaptive visual update minimizes visual features added to the map when LiDAR constraints are sufficient, further lowering the memory usage.
3) Extensive experiments validation and dataset open-source: Validation was performed on both public and private datasets, with the private datasets featuring more challenging and aggressive scenarios. Experiments were conducted on x86 personal laptops and cost-effective ARM devices, demonstrating superior efficiency. Private Datasets and supplementary materials, including further method details and additional experiments are released $^{1}$.
### II. RELATED WORKS
The topic of this work is most relevant to degeneration-aware strategy and map structure in odometry system, which are discussed as follows.
### A. Degeneracy-Aware Odometry Systems
Existing studies have explored LiDAR odometry degeneracy and developed adaptive strategies within odometry systems, leveraging updated degeneration states to enhance robustness in challenging environments. Zhang et al. [^10] evaluated degeneracy of optimization-based methods through analysis of geometric structure. Tuna et al. [^11] proposed X-ICP, integrating localizability detection and localizability aware optimization based on the classical mechanics of a point cloud. AdaLIO [^12] adaptively adjust some key parameters after checking degeneracy. Hatleskog et al. [^13] proposed a probabilistic method for LiDAR degeneracy detection, and incorporated it into a degeneracy-aware ICP framework. Zhu et al. [^14] proposed a degeneration evaluation method to dynamically adjust the state vector to ensure robust and efficient swarm state estimation. Lee et al. [^15] implemented a non-heuristic degeneracy detection using a predefined threshold on normalized eigenvalues and further use it to select a better initial guess in a switch structure. More recently, GenZ-ICP [^16] combines point-to-point and point-to-plane metrics via adaptive weighting, enhancing robustness in geometrically degenerate scenarios like corridors.
While previous works primarily focus on odometry robustness, our system harnesses the potential of degeneration-aware factors to dynamically adjust computational burden allocation based on environmental constraints, achieving both robustness and notable computational efficiency. Our design embodies the principle of passive-degeneracy mitigation [^17], well-suited for the mildly to severely degenerate environments we target.
### B. Map Structure in LiDAR-Visual-(Inertial) Odometry
In LiDAR-visual-(inertial) odometry, the map structure is a critical factor for determining both efficiency and accuracy. Generally, map structures are categorized into three types. The
Fig. 1. Overview of a real-time experiment on the low-power ARM platform. (a) The ARM platform RK3588 with an octa-core architecture (4× Cortex-A76 + 4× Cortex-A55) and a maximum frequency of 2.4 GHz used for testing, priced at approximately 100 USD. (b) Point cloud of the nighttime street scene used for testing, with the orange trajectory representing the collected path. (c) Detailed per-frame runtime statistics of the system, with data input from both LiDAR and camera at 10 Hz.
first type maintains two separate maps for the visual and LiDAR components, where the visual backend generates 3D points independently. In the second type, two maps are still maintained, but the visual backend does not generate 3D points independently. The third type employs a unified map for both visual and LiDAR measurements.
For instance, in LVI-SAM [^3] and R2LIVE [^6], VINS-mono [^18] serves as the visual subsystem, which falls under the first category. In the second category, systems like SDVLOAM [^19], DV-LOAM [^20], FAST-LIVO [^4], PL-LIVO [^21], and CamVox [^22] reuse LiDAR points to provide depth information for visual features, patches, or line patches. However, these systems still store visual and LiDAR map points in two distinct data structures. Notably, SDV-LOAM and DV-LOAM limit the use of visual map points to keyframes or the last frame, whereas FAST-LIVO and PL-LIVO incorporate visual map points from the global map that fall within the current field of view (FoV).
In contrast, the third category includes SR-LIVO [^23], R3LIVE [^7], FAST-LIVO2 [^5], and our system, which all maintain a single unified map. SR-LIVO and R3LIVE adopt an incremental k-d tree (ikd-tree [^24]) to manage point-based maps, tagging each point with RGB color. Meanwhile, FAST-LIVO2 employs a hash-indexed octree to store a surfel-based map, where each leaf node contains a LiDAR plane feature augmented with an image patch. Our system further advances this approach by maintaining a compact unified local map and preserving a lightweight long-term visual map, enabling robust odometry performance with considerably reduced memory consumption.
### III. SYSTEM OVERVIEW
The system overview is shown in Fig. 2. At measurement level, as the input of this system, high-frequency raw LiDAR points are segmented into distinct scans employing scan recombination. To optimize computational efficiency, an adaptive visual frame selector dynamically selects images based on environmental constraints, allowing our degeneration-aware odometry to allocate computational resources adaptively. For
---
Fig. 2. System overview. In the figure detailing voxel data on the lower-right side, the dashed ellipse encloses points considered to lie on a plane with varying scales.
Fig. 3. Illustration of sequential update ESIKF with adaptive visual frame selector.
state estimation, we construct specific residuals, including LiDAR point-to-plane residuals and photometric errors, and tightly couple all sensor measurements through an ESIKF with sequential updates. At the mapping level, a robocentric, compact unified local map and a lightweight long-term visual map are maintained through efficient integration of new observations, sliding operations, and visual features rearrangement, ensuring memory efficiency and robust performance.
### IV. METHODOLOGY
Previous work FAST-LIVO2 [^5] established the core framework for multi-sensor (LiDAR-inertial-visual) integration, including state definitions, discrete transition models, LiDAR-visual map management, and state estimation by fusing camera and LiDAR measurements. Expanding on this foundation, this work focuses on building a lightweight LIVO system deploying on resource-constraint platforms. Therefore, in this section, we emphasize our contributions on achieving computationally efficient state estimation (Section IV-A) and memory-saving map (Section IV-B).
### A. State Estimation With Visual Frame Selector
We first introduce our LiDAR degeneration evaluation method (Section IV-A1), followed by the adaptive visual frame selector (Section IV-A2) enabling more efficient ESIKF-based state estimation, as depicted in Fig. 3.
1) LiDAR Degeneration Evaluation: After obtaining N undistorted points $^L\mathbf{p} \in \mathbf{R}^{3 \times N}$ from one recombined LiDAR scan, where each point is assumed to lie exactly on a plane, the point-to-plane LiDAR measurement model can be expressed as follows:
$$ \mathbf{0}=\left(\mathbf{n}^{g t}\right)^{T}\left(^{G}\mathbf{T}_{L}\left(^{L}\mathbf{p}-\delta^{L}\mathbf{p}\right)-\mathbf{q}^{g t}\right), $$
where $^G\mathbf{T}_L$ denotes the pose of the LiDAR frame with respect to the global frame. The term $\delta^L\mathbf{p}$ accounts for the beam noise between the measured $^L\mathbf{p}$ and its ground truth. Ideal corresponding planes are parameterized by normal vector $\mathbf{n}^{gt} \in \mathbf{R}^{3 \times N}$ and point on the plane $\mathbf{q}^{gt} \in \mathbf{R}^{3 \times N}$. Additionally, $\mathbf{n}^{est}$ and $\mathbf{q}^{est}$ are the estimated plane features used in the subsequent LiDAR degeneration analysis.
In the context of LiDAR degeneration evaluation, it is exceedingly rare for all three translational degrees of freedom (DoFs) to remain well-constrained while the three rotational DoFs experience degeneration (e.g., within an ideal spherical structure). Based on this observation, we simplify the problem by focusing solely on the translational constraints.
State estimation is performed by minimizing the residuals from (1) as follows:
$$ \arg\min_{^{G}\mathbf{R}_{L},^{G}\mathbf{t}_{L}}\left\|\underbrace{(\mathbf{n}^{est})^{T}}_{\mathbf{A}}\underbrace{^{G}\mathbf{t}_{L}}_{\mathbf{x}}+\underbrace{(\mathbf{n}^{est})^{T}(^{G}\mathbf{R}_{L}{}^{L}\mathbf{p}-\mathbf{q})}_{\mathbf{b}}\right\|^{2}. $$
The degeneration of LiDAR can be analyzed by evaluating the stability of this associated optimization problem. The objective function, expressed as $\|\mathbf{A}\mathbf{x} - \mathbf{b}\|^2$, can be evaluated the solution stability by analyzing the singular values of $\mathbf{A}^T\mathbf{A}$, as inspired by [^10]. To enhance generalizability across different LiDAR types and parameter settings (e.g., LiDAR scan downsampling strategies), the singular value vector is further normalized as follows:
$$ \left[\tilde{\sigma}_{\min},\tilde{\sigma}_{\mathrm{m i d}},\tilde{\sigma}_{\max}\right]=\mathrm{N o r m a l i z e}\left(\mathrm{S V D}\left(\mathbf{n}^{\mathrm{e s t}}\left(\mathbf{n}^{\mathrm{e s t}}\right)^{T}\right)\right) $$
The smallest normalized value, $\tilde{\sigma}_{\mathrm{min}}$ is compared against a predefined threshold. If $\tilde{\sigma}_{\mathrm{min}}$ remains below this threshold for multiple consecutive frames, the LiDAR module is considered to be in a degenerate state; otherwise, the environment is deemed to provide sufficient constraints for LiDAR scans.
---
2) Adaptive Visual Frame Selector: As shown in Fig. 2, we achieve an adaptive image selector based on the degeneration state of the current LiDAR scans and the motion of the sensor. Specifically, when LiDAR degeneration occurs (e.g., when small-FoV AVIA LiDAR faces a wall), all available images are utilized to construct as many constraints as possible, preventing localization failure, as the camera provides robust constraints that contribute to system robustness in most cases. Under normal conditions, only sparsely distributed keyframes are selected for updating state and visual related observation in map at a reduced frequency, minimizing computational costs.
Keyframes are defined as images captured when the sensor's pose (position or orientation) changes markedly, exceeding a threshold $\tau$ relative to the previous frame. This threshold is adaptively adjusted as:
$$ \tau=\sqrt{3}\cdot\tilde{\sigma}_{\mathrm{m i n}}\cdot\tau_{\mathrm{p r e d f e n e d}}, $$
where $\tau_{\text{predefined}} \in \mathbb{R}_+$ contain predefined thresholds for position and orientation. The factor $\sqrt{3} \cdot \tilde{\sigma}_{\min} \in [0,1]$ scales the visual keyframe selection threshold adaptively since $\tilde{\sigma}_{\min}$ reflects the sufficiency of environmental localization constraints from LiDAR, thus contributing to system robustness even before full degeneration occurs.
### B. Map Management
1) Map Structure: The entire map is divided into a unified local map and a long-term visual map. As depicted in the right part of Fig. 2, the map, consist of adaptive voxels, is managed by one hash table. Specifically, the Hash value here is a global coordinate index of the root voxel centre LOCATION ∈ N³ for easier retrieval. The fixed size of root voxel is set to 0.5 × 0.5 × 0.5 meters.
Unified Local Map contains dense points and plane feature (i.e., plane center, normal vector and associated uncertainty) at various scales within a three-level octree structure. Some points selected as visual points are attached with a three-level pyramid of image patches for visual state estimation, while the remaining points retain only geometric information for LiDAR scan-to-map matching and subsequent LiDAR state estimation.
Long-term Visual Map is a collection of historical points with visual observations for visual frame-to-map match. This map is relatively sparse in space, allowing long-term environmental observations to be stored with relatively small memory consumption, while providing essential historical data for long-term localization.
2) Map Updating and Sliding: The process of registering newly observations and estimated plane features to the unified local map follows the approach outlined in [^5], [^25]. Shown in Fig. 4, instead of shifting the large-sized local map only when the sensor range reaches the map boundary, as done in FAST-LIVO2, we adopt a relatively small-sized unified local map with a more frequent map-sliding strategy. Specifically, when the robot's positional movement surpasses a predefined threshold since the last map sliding, plane features and point clouds outside the local map boundary are cleared to maintain efficiency. Points with visual observations that fall outside the small unified local map are transferred to the long-term visual map. The long-term visual map follows a similar sliding mechanism but operates on a larger scale to handle the extended range.
Fig. 4. Illustration of map structure and sliding process in our system and FAST-LIVO2.
Fig. 5. Our platform with hardware synchronization for data acquisition.
### V. EXPERIMENT AND RESULTS
In this section, we first demonstrate the overall performance of our system on public and private datasets, followed by module-specific evaluations.
### A. Setup of the Experiments
1) Public Dataset: We evaluate our method using the Hilti datasets [^26], [^27], which feature indoor and outdoor sequences captured in challenging environments such as construction sites, offices, and basements. These datasets include handheld and robot-mounted configurations with LiDAR, cameras, and IMUs operating at different frequencies.
2) Private Dataset: We validated the system's robustness using a B/W camera with a fisheye lens and a Livox Mid-360 LiDAR, synchronized via an STM32 microcontroller. In extremely dark environments, a 15 W onboard illuminator was used to enhance visibility Fig. 5.
3) Experimental Settings and Configurations: The experiments were conducted on two hardware platforms, referred to as x86 and ARM platforms for simplicity. The x86 platform is a personal laptop with a 13th Gen Intel Core i9-13900HX CPU. The ARM platform, priced at approximately $100, is notably more cost-effective and features an RK3588 processor with an octa-core architecture (4× Cortex-A76 + 4× Cortex-A55) and a maximum frequency of 2.4 GHz.
A consistent degeneration detection threshold of 0.07 was used across all tests. The local unified map and the long-term visual map were configured with typical edge lengths of 200 m and 800 m, uniformly across all sequences, with map sliding thresholds of 20 m and 100 m of robot motion, respectively. Additionally, the keyframe selection thresholds were adapted to the scene scale, with values of 1m-60° for all indoor sequences.
---
Fig. 6. Visualization of some representative tested sequences. $*$ denotes the moment when detected LiDAR degeneration occurs. The percentage next to each sequence name indicates the proportion of selected images relative to all available images in that sequence.
Fig. 7. Real-time localization experiments on the RK3588 platform. (a) Underground parking lot and (b) nighttime street. For each scenario, the left image shows the trajectory and point cloud, and the right image shows a third-person view at a specific moment.
and 2m-60° for all outdoor sequences. The equidistant projection model was employed to handle fisheye camera images in our private dataset. Other parameters, such as noise settings for different LiDAR models, were inherited from FAST-LIVO2, which had been carefully tuned for robust performance.
### B. System Performance on Public Datasets
1) Accuracy on Public Datasets: In this experiment, we first validated the accuracy of our method on 16 sequences from the Hilti'22 and Hilti'23 datasets, benchmarking it against
Fig. 8. Mean LiDAR and visual processing times for our method and FAST-LIVO2 across seven scenarios. LiDAR mean time forms the base, with visual mean time stacked above. Error bars indicate the observed range.
state-of-the-art systems, including R3LIVE [^7], FAST-LIO2 [^2], SDV-LOAM [^19], LVI-SAM [^3], and our previous work, FAST-LIVO series [^4], [^5]. We made efforts to tune the parameters of all methods to achieve their best performance for comparison. Using Hilti official evaluation tools, $^{2}$ the RMSE results are shown in Table I. From the experimental results, our system
---
TABLE I ATE (RMSE, M) FOR HILTI DATASETS
| Dataset | Sequence | Ours | FAST-LIVO2 | FAST-LIO2 | FAST-LIVO | R3LIVE | SDV-LOAM | LVI-SAM |
| Hiti '22 | Construction Ground | 0.010 | 0.010 | 0.013 | 0.022 | 0.021 | 25.121 | ✗ |
| Construction Multilevel | 0.023 | 0.020 | 0.044 | 0.052 | 0.024 | 12.561 | ✗ |
| Construction Stairs | 0.170 | 0.016 | 0.320 | 0.241 | 0.784 | 9.212 | 9.142 |
| Long Corridor | 0.054 | 0.067 | 0.064 | 0.065 | 0.061 | 19.531 | 6.312 |
| Cupola | 0.220 | 0.121 | 0.250 | 0.182 | 2.142 | 9.321 | ✗ |
| Lower Gallery | 0.018 | 0.007 | 0.024 | 0.022 | 0.008 | 11.232 | 2.281 |
| Attic to Upper Gallery | 0.180 | 0.069 | 0.720 | 0.621 | 2.412 | 4.551 | ✗ |
| Outside Building | 0.041 | 0.035 | 0.028 | 0.052 | 0.029 | 2.622 | 0.952 |
| Hiti '23 | Floor 0 | 0.032 | 0.021 | 0.031 | 0.022 | 0.024 | 4.621 | ✗ |
| Floor 1 | 0.018 | 0.023 | 0.031 | 0.022 | 0.024 | 7.951 | 8.682 |
| Floor 2 | 0.038 | 0.022 | 0.083 | 0.048 | 0.046 | 7.912 | ✗ |
| Basement | 0.024 | 0.016 | 0.038 | 0.035 | 0.024 | 6.151 | ✗ |
| Stairs | 0.012 | 0.018 | 0.170 | 0.152 | 0.110 | 9.032 | 3.584 |
| Parking 3x Floors Down | 0.095 | 0.032 | 0.320 | 0.356 | 0.462 | 19.952 | ✗ |
| Large room | 0.028 | 0.026 | 0.028 | 0.031 | 0.035 | 16.781 | 0.563 |
| Large room (dark) | 0.044 | 0.046 | 0.040 | 0.053 | 0.059 | 15.012 | ✗ |
| Average | | 0.063 | 0.034 | 0.138 | 0.123 | 0.391 | 11.347 | 4.502 |
$^{*}$ denotes the system totally failed.
—Underline denotes the second-best, next to the bolded result.
achieved the second-highest average RMSE accuracy of 6.3 cm, surpassed only by FAST-LIVO2's 3.4 cm, while outperforming the state-of-the-art LiDAR-inertial only method FAST-LIO2. The integration of the visual module within our system proves beneficial, as it enhances localization accuracy.
Specifically, in sequences with limited visual texture (e.g., Long Corridor and Stairs), our system outperformed FAST-LIVO2 by selectively reducing visual usage when LiDAR constraints were robust. However, in visually challenging scenarios (e.g., the overexposed Outside Building or the poorly lit Large Room (dark)), the system exhibited slightly reduced accuracy compared to FAST-LIO2. On the Hiti datasets, SDV-LOAM demonstrated substantial limitations, primarily due to the absence of tight IMU integration, which caused notable drift within its LiDAR Odometry subsystem. Additionally, the system's loose coupling between LiDAR and visual measurements, along with insufficient initialization of Visual Odometry, frequently led to convergence issues, such as local optima or failed optimizations. LVI-SAM experienced failures in nine sequences, as its feature-based LIO and VIO subsystems did not fully utilize raw measurements, making it less robust in environments with subtle geometric or texture features (e.g., minimally textured or geometrically simple scenes). While R3LIVE achieved generally strong results, its performance deteriorated in sequences involving intense rotations and sparse structural information (e.g., Construction Stairs, Cupola and Attic to Upper Gallery). In these scenarios, inadequate pose priors during map alignment caused convergence to suboptimal solutions, ultimately resulting in optimization failures.
2) Computation Efficiency on Public Datasets: We evaluated the algorithm’s runtime on 16 sequences from the Hilti dataset using both x86 and ARM platforms, $^{3}$ as summarized in Table II. Generally, our system achieved a notable improved computational efficiency compared to FAST-LIVO2. This improvement is primarily attributed to the adaptive frame rate strategy in the visual module, which significantly reduced the runtime of visual processing. Additionally, the LiDAR module experienced a slight increase in runtime due to optimizations in the code implementation.
TABLE II
THE MEAN AND STANDARD DEVIATION OF TIME CONSUMPTION IN HILTI SEQUENCES (MSEC)
| Architecture | ARM | x86 PC |
| SLAM System | Ours | FAST-LIVO2 | Ours | FAST-LIVO2 |
| LiDAR Part | 53.83 / 5.35 | 56.50 / 5.56 | 23.36 / 2.36 | 25.05 / 2.41 |
| Visual Part | 3.99 / 1.80 | 19.36 / 4.75 | 2.64 / 1.12 | 10.61 / 2.68 |
| Total | 57.82 / 6.75 | 75.87 / 9.77 | 25.99 / 3.22 | 35.66 / 4.88 |
TABLE III
MEMORY USAGE COMPARISON BETWEEN OUR ALGORITHM AND FAST-LIVO2 FOR HILTI DATASET SEQUENCES (GB)
| Dataset | Sequence | Ours | FAST-LIVO2 |
| Hiti '22 | Construction Ground | 3.0 | 4.1 |
| Construction Multilevel | 3.7 | 4.9 |
| Construction Stairs | 1.3 | 2.4 |
| Long Corridor | 0.8 | 1.2 |
| Cupola | 1.9 | 3.4 |
| Lower Gallery | 1.0 | 1.6 |
| Attic to Upper Gallery | 1.4 | 2.4 |
| Outside Building | 2.4 | 3.0 |
| Hiti '23 | Floor 0 | 2.4 | 3.1 |
| Floor 1 | 1.5 | 2.2 |
| Floor 2 | 1.7 | 2.3 |
| Basement | 1.3 | 2.4 |
| Stairs | 1.3 | 1.8 |
| Parking 3x Floors Down | 1.7 | 2.6 |
| Large room | 1.0 | 1.6 |
| Large room (dark) | 1.1 | 1.6 |
| Average | | 1.7 | 2.5 |
3) Memory Consumption on Public Datasets: We further evaluate our method's computational efficiency and memory consumption on the public dataset. Since FAST-LIVO2 demonstrated the highest accuracy among all compared methods, as shown in Table I, we focus on a comparison between it with our proposed system to highlight the improvements in resource efficiency. We measure the memory usage of the validated Hiti sequences on x86 platform, summarized in the table III. It can be observed that our system notably reduces memory consumption compared to FAST-LIVO2. Since the algorithm's memory consumption is primarily attributed to map storage, the reduction in small-scale scenarios is largely due to the adaptive image frame selector, which effectively minimizes redundant visual observations in the map. For larger-scale scenarios, the
---
LiDAR Degeneration Evaluation on AVIA
LiDAR Degeneration Evaluation on MID-360
Fig. 9. Validation of LiDAR degeneracy evaluation. The plot shows the smallest normalized singular value, with degeneration detected when the value falls below the threshold (dashed line), indicated by red shaded areas. The top visualizations correspond to moments of low $(t_{1}, t_{3})$ and high $(t_{2}, t_{4})$ singular values, with red points representing the current LiDAR scan and the top-right image showing the corresponding first-person camera view.
Fig. 10. Ablation study results on our proposed long-term visual map structure, presented as reconstructed colored point clouds. (a) With the proposed long-term visual map. (b) Without the long-term visual map.
Fig. 11. Comparison of memory usage for different algorithms in MaRS LVIG sequence HKIsland03.
memory reduction is primarily attributed to our designed map structure, which utilizes a smaller-sized local unified map to efficiently manage storage.
### C. System Performance on Private Datasets
We conducted qualitative tests on the system's robustness and computational real-time performance using private datasets.
1) Experiments on Challenging Private Sequences: We conducted tests on challenging private sequences. Illustrated in Fig. 6, these sequences encompass a variety of SLAM challenges, such as dynamic lighting conditions in Mining Tunnel, extremely low illumination in Dark Woods, unstructured environments in Wild Park and severe LiDAR degeneration in HIT Graffiti Wall. Additionally, they cover both indoor and outdoor scenes captured at different times, also including indoor scenarios with lights turned off (e.g., HW Corridor). As an experimental result, our system demonstrated robust performance, producing sharp point cloud maps and even vivid colourful point clouds. In all sequences where the data collection equipment physically returns to the starting position, our system achieves low return-to-origin drift of less than 2 cm.
Meanwhile, the percentage in Fig. 6 refers to the proportion of selected images relative to all available images, showing the system's adaptive sensitivity to the environmental constraints provided by LiDAR. For instance, in long tunnel-like corridors such as the Mining Tunnel and mid part of HW Corridor, dense visual frames are selected due to weak LiDAR constraints. In contrast, in environments with strong LiDAR constraints, such as the HKUST Red Sculpture, images are selected more sparsely in space.
2) Real-Time Localization Experiments on Resource-Constrained ARM Platform: Shown in Fig. 7, we conducted real-time localization tests in the evening on bustling streets with pedestrian and in an underground parking lot with vehicles exiting, as well as substantial indoor-outdoor lighting variations. All computations were performed entirely on the onboard ARM platform. The results demonstrated robust localization with a per-frame time cost of 37 ms.
### D. Evaluation of Adaptive Image Frame Selector
1) Evaluation of Computational Efficiency Optimization Via Adaptive Image Frame Selector: From the visualization results of the selected visual frames in Fig. 6, it can be observed that different proportions of images were selected from all available visual measurements across various sequences, based on the LiDAR degeneration conditions in the environment. We analyzed the LiDAR and visual computation times and compared them with FAST-LIVO2. As shown in Fig. 8, our method achieves substantially lower average visual computation times, with comparable maximum but notably lower minimum per-iteration costs, especially in sequences with strong LiDAR constraints where visual overhead is minimal.
2) Evaluation of LiDAR Degeneration Detection Module: As illustrated in Fig. 9, we conducted a detailed analysis of the effectiveness of the LiDAR degeneration detection module using two highly challenging sequences for LiDAR. The results highlight the algorithm's sensitivity in identifying degeneration, as evidenced by extremely low eigenvalues in degenerate scenarios (e.g., the Aivia LiDAR facing a wall at $t_1$ and the Mid360 LiDAR in a long, straight tunnel at $t_3$). Furthermore, the analysis
---
confirms that a universal degeneration threshold can be applied across different LiDAR types, demonstrating the effectiveness of our LiDAR degeneration evaluation module.
### E. Evaluation of Map Structure
To evaluate the proposed long-term visual map, we perform an ablation study to assess its impact on localization accuracy and memory usage using the large-scale sequence HKIsland03 in MARS-LVIG dataset [^28].
1) Accuracy Comparison: The localization accuracy is evaluated using the RTK trajectory as ground truth. We evaluated RMSE using the evo toolkit. $^{4}$ In this sequence, the long-term visual map improved accuracy, achieving better RMSE 0.39 m related to 0.85 m, as evidenced by the detailed point cloud visualization in Fig. 10.
2) Memory Consumption Comparison: Shown in Fig. 11, while the long-term visual map introduces additional memory overhead, its impact on memory usage is relatively modest compared to the significant improvement in localization accuracy. This trade-off is justified, as the enhanced robustness and accuracy, confirming the map's suitability for large-scale deployments. Meanwhile, the unified local map edge length parameter of FAST-LIVO2 was set to 2 km, as in the original paper. With the scenario contained within the map boundaries, memory usage continuously increases.
### VI. CONCLUSION
This paper presents a lightweight LIVO system optimized for resource-constrained platforms. By integrating a degeneration-aware adaptive visual frame selector and a memory-efficient mapping structure, the system significantly reduces computation and memory usage while maintaining odometry accuracy. Extensive experiments validate its robust performance and demonstrate its capability to run efficiently on resource-constrained edge devices, enabling real-time operation in challenging environments.
## REFERENCES
[^1]: J. Zhang and S. Singh, "LOAM: LiDAR odometry and mapping in real-time," Robot. Sci. Syst., vol. 2, no. 9, pp. 1–9, 2014.
[^2]: W. Xu and F. Zhang, “FAST-LIO: A fast, robust LiDAR-inertial odometry package by tightly-coupled iterated Kalman filter,” IEEE Robot. Automat. Lett., vol. 6, no. 2, pp. 3317–3324, Apr. 2021.
[^3]: T. Shan, B. Englot, C. Ratti, and D. Rus, “LVI-SAM: Tightly-coupled LiDAR-visual-inertial odometry via smoothing and mapping,” in Proc. IEEE Int. Conf. Robot. Automat., Xi’an, China, 2021, pp. 5692–5698.
[^4]: C. Zheng, Q. Zhu, W. Xu, X. Liu, Q. Guo, and F. Zhang, "FAST-LIVO: Fast and tightly-coupled sparse-direct LiDAR-inertial-visual odometry," in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., Kyoto, Japan, 2022, pp. 4003–4009.
[^5]: C. Zheng et al., “FAST-LIVO2: Fast, direct LiDAR-inertial-visual odometry,” IEEE Trans. Robot., pp. 326–346, 2024.
[^6]: J. Lin, C. Zheng, W. Xu, and F. Zhang, “R $^{2}$ LIVE: A robust, real-time, LiDAR-inertial-visual tightly-coupled state estimator and mapping,” IEEE Robot. Automat. Lett., vol. 6, no. 4, pp. 7469–7476, Oct. 2021.
[^7]: J. Lin and F. Zhang, “R $^{3}$ LIVE: A robust, real-time, RGB-colored, LiDAR-inertial-visual tightly-coupled state estimation and mapping package,” in 2022 Int. Conf. Robot. Automat., Philadelphia, PA, USA, 2022, pp. 10672–10678.
[^8]: X. Zhou, Z. Wang, H. Ye, C. Xu, and F. Gao, “EGO-planner: An ESDF-free gradient-based local planner for quadrotors,” IEEE Robot. Automat. Lett., vol. 6, no. 2, pp. 478–485, Apr. 2020.
[^9]: G. Lu, W. Xu, and F. Zhang, “On-manifold model predictive control for trajectory tracking on robotic systems,” IEEE Trans. Ind. Electron., vol. 70, no. 9, pp. 9192–9202, Sep. 2023.
[^10]: J. Zhang, M. Kaess, and S. Singh, “On degeneracy of optimization-based state estimation problems,” in Proc. IEEE Int. Conf. Robot. Automat., Stockholm, Sweden, 2016, pp. 809–816.
[^11]: T. Tuna, J. Nubert, Y. Nava, S. Khattak, and M. Hutter, “X-ICP: Localizability-aware LiDAR registration for robust localization in extreme environments,” IEEE Trans. Robot., vol. 40, pp. 452–471, 2024.
[^12]: H. Lim, D. Kim, B. Kim, and H. Myung, “AdaLIO: Robust adaptive LiDAR-inertial odometry in degenerate indoor environments,” in Proc. 20th Int. Conf. Ubiquitous Robots, Honolulu, HI, USA, 2023, pp. 48–53.
[^13]: J. Hatleskog and K. Alexis, “Probabilistic degeneracy detection for point-to-plane error minimization,” IEEE Robot. Automat. Lett., vol. 9, no. 12, pp. 11234–11241, Dec. 2024.
[^14]: F. Zhu et al., “Swarm-LIO2: Decentralized, efficient LiDAR-inertial odometry for UAV swarms,” IEEE Trans. Robot., vol. 41, pp. 960–981, 2025.
[^15]: J. Lee et al., “Switch-SLAM: Switching-based LiDAR-inertial-visual SLAM for degenerate environments,” IEEE Robot. Automat. Lett., vol. 9, no. 8, pp. 7270–7277, Aug. 2024.
[^16]: D. Lee, H. Lim, and S. Han, “GenZ-ICP: Generalizable and degeneracy-robust LiDAR odometry using an adaptive weighting,” IEEE Robot. Automat. Lett., vol. 10, no. 1, pp. 152–159, Jan. 2025.
[^17]: T. Tuna, J. Nubert, P. Pfreundschuh, C. Cadena, S. Khattak, and M. Hutter, “Informed, constrained, aligned: A field analysis on degeneracy-aware point cloud registration in the wild,” 2024, arXiv:2408.11809.
[^18]: T. Qin, P. Li, and S. Shen, “VINS-mono: A robust and versatile monocular visual-inertial state estimator,” IEEE Trans. Robot., vol. 34, no. 4, pp. 1004–1020, Aug. 2018.
[^19]: Z. Yuan, Q. Wang, K. Cheng, T. Hao, and X. Yang, “SDV-LOAM: Semi-direct visual–LiDAR odometry and mapping,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 45, no. 9, pp. 11203–11220, Sep. 2023.
[^20]: W. Wang, J. Liu, C. Wang, B. Luo, and C. Zhang, "DV-LOAM: Direct visual LiDAR odometry and mapping," Remote Sens., vol. 13, no. 16, 2021, Art. no. 3340.
[^21]: T. Shi, K. Qian, Y. Fang, Y. Zhang, and H. Yu, “Point-line LIVO using patch-based gradient optimization for degenerate scenes,” IEEE Robot. Automat. Lett., vol. 9, no. 11, pp. 9717–9724, Nov. 2024.
[^22]: Y. Zhu, C. Zheng, C. Yuan, X. Huang, and X. Hong, “CamVox: A low-cost and accurate lidar-assisted visual SLAM system,” in Proc. IEEE Int. Conf. Robot. Automat., Xi’an, China, 2021, pp. 5049–5055.
[^23]: Z. Yuan, J. Deng, R. Ming, F. Lang, and X. Yang, “SR-LIVO: LiDAR-inertial-visual odometry and mapping with sweep reconstruction,” IEEE Robot. Automat. Lett., vol. 9, no. 6, pp. 5110–5117, Jun. 2024.
[^24]: Y. Cai, W. Xu, and F. Zhang, "ikd-tree: An incremental K-D tree for robotic applications," 2021, arXiv:2102.10808.
[^25]: C. Yuan, W. Xu, X. Liu, X. Hong, and F. Zhang, “Efficient and probabilistic adaptive voxel mapping for accurate online LiDAR odometry,” IEEE Robot. Automat. Lett., vol. 7, no. 3, pp. 8518–8525, Jul. 2022.
[^26]: L. Zhang et al., “Hilti-Oxford dataset: A millimeter-accurate benchmark for simultaneous localization and mapping,” IEEE Robot. Automat. Lett., vol. 8, no. 1, pp. 408–415, Jan. 2023.
[^27]: M. Helmberger, K. Morin, B. Berner, N. Kumar, G. Cioffi, and D. Scaramuzza, “The Hilti SLAM challenge dataset,” IEEE Robot. Automat. Lett., vol. 7, no. 3, pp. 7518–7525, Jul. 2022.
[^28]: H. Li et al., “MARS-LVIG dataset: A multi-sensor aerial robots SLAM dataset for LiDAR-visual-inertial-GNSS fusion,” Int. J. Robot. Res., vol. 43, no. 8, pp. 1114–1127, Jan. 2024.