Enhanced Isolation Forest-Based Algorithm for Unsupervised Anomaly Detection in Lidar SLAM Localization
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe paper is well organized. Please carefully address the following questions/comments:
1) What are the main causes of drift or localization failures in SLAM systems that the paper addresses?
2) Why is it important to introduce the Isolation Forest algorithm in detecting localization anomalies? In other words, what are the merits of the proposed approach with respect to these papers: 10.1109/AIIoT54504.2022.9817309; 10.1109/IGESSC55810.2022.9955337. Please clearly explain.
3) How does the path weighting method enhance the scoring mechanism in the proposed model?
4) What role does the adaptive OTSU algorithm play in improving the model's reliability?
5) Why do the experimental results demonstrate that the proposed model outperforms other SOTA algorithms in terms of sensitivity?
Author Response
Thank you for your detailed reading and effective comments. Please check the attachment for our modification.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe paper is devoted to the method of detecting anomalies in localization.
Introduction
The paper is devoted to the method of detecting anomalies in localization using the Isolation Forest (iForest) algorithm and the adaptive threshold value determined by the Otsu method. The study is conducted based on the KITTI dataset, which is classified into several categories of environments, such as urban environments with complex roads, short trajectories, direct circular route, etc.
Methodology
1. Classic Isolation Forest (iForest)
- Construction of an isolation tree (iTree): Subsamples are randomly selected from the original data, from which t isolation trees are built. For each data point, iForest recursively uses random features and random split values ​​to divide the data, forming a binary tree (iTree). This process continues until only one data point remains in each subspace.
- Path length calculation: The path length (hx) for a data point x is defined as the number of edges traversed from the root node to a leaf node in an iTree. This path length indicates how deeply isolated the data point is.
- Average path length calculation: For a given data point x, the average path length (E[h(x)]) is calculated as the average of the path lengths across all iTrees in the iForest.
- Path length normalization: The average path length E[h(x)] is normalized using a reference value c(n), which is derived from the harmonic number approximation. This scales the path length, ensuring that the anomaly score s(x,n) is the same regardless of sample size.
- Anomaly score calculation: The anomaly score s(x,n) for a data point x is calculated using formula (18). Scores closer to 1 indicate a higher probability of an anomaly.
2. Improved anomaly scoring formula
- The classic iForest anomaly scoring formula only takes into account the path length, assuming that all nodes in the isolation trees have the same importance. This is a rough assumption. In this paper, we introduce a path-weighted score that encodes the weight of each traversed node, and this weight is related to the structure of the isolation trees.
- A new feature (t hx) is used to extract the path length x in tree t, expressed by formula (19). The weight tk W is calculated by formula (20), where tk N represents the number of samples contained in node k of isolation tree t.
- The improved anomaly scoring formula is represented by formula (21). By optimizing the formula, higher weights are assigned to anomalous samples (those that split earlier), while the weights for normal samples are reduced. These weights are determined based on the structure of the isolation tree, namely the number of samples contained in node k. The improved scoring formula shows greater sensitivity to local anomalies.
3. Adaptive thresholding using Otsu's method
- In order to make the algorithm proposed in this paper closer to a fully unsupervised method, the Otsu thresholding method is used to adaptively set the anomaly scoring threshold. In the classical iForest algorithm, the anomaly scoring threshold is set manually, usually based on the anomaly scoring contour, which is a rather fuzzy method. The threshold has a significant impact on the accuracy of anomaly detection, so the selection of the scoring threshold is especially important. In this paper, the Otsu method is used to accurately determine the anomaly scoring threshold.
Main Results
1. Accuracy and Sensitivity:
- The algorithm shows high accuracy in anomaly detection. It can simultaneously detect changes in each pose parameter (translation and rotation) and cover all aspects of anomaly recognition.
- The results are within the standard deviation (Std) of the EVO (Evaluation of Visual Odometry) calculations and close to the mean (Mean) results.
2. Comparison with other methods:
- The algorithm showed superiority over ablative methods such as IF-OTSU, IF, K-means and SVM.
- Compared with the current standard (SOTA) methods, the proposed algorithm is more sensitive to anomalies.
3. Timing characteristics:
- Algorithm execution time:
- Maximum time: 0.027024 seconds
- Minimum time: 0.0094382 seconds
- Average time: 0.017034467 seconds
- These timing characteristics indicate the high efficiency of the algorithm, which allows it to be used in real time.
Conclusion
The paper demonstrates that the proposed anomaly detection algorithm in pose estimation LIO-SAM is effective and sensitive. It can be used in real time to monitor and analyze variables, making it versatile and applicable to various systems. The results show that the algorithm is not limited to a specific structure and can be adapted to various tasks.
Recommendations
- Further Research:
- Further research is recommended to evaluate the performance of the algorithm on more diverse datasets and under different conditions.
- Practical Applications:
- The algorithm can be integrated into autonomous vehicle control systems, robotics, and other areas where accurate and reliable pose estimation is required.
This paper represents a significant contribution to the field of anomaly detection in pose estimation and can serve as a basis for further research and development.
Notes: Page 7 has locus, better to add text in the end of page 7.Author Response
Thank you for your detailed reading and effective comments. Please check the attachment for our modification.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for Authors1- The abstract is well-structured and provides a clear motivation for the study. However, adding a brief statement about the experimental setup (e.g., dataset, number of test cases) would improve clarity.
2-It would be beneficial to elaborate on how the proposed model handles dynamic obstacles and environmental changes in real-world scenarios.
3-The paper could discuss potential trade-offs between sensitivity and false positives in anomaly detection
4- conterbution of paper should explain explacitly
5-How is variable correlation analysis performed, and how does it impact the model's feature selection process?
6-What specific advantages does the path-weighting mechanism provide over traditional Isolation Forest approaches?
Comments on the Quality of English Languageenglish language can be improve
Author Response
Thank you for your detailed reading and effective comments. Please check the attachment for our modification.
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for AuthorsThe article “Enhanced Isolation Forest-Based Algorithm for Unsupervised Anomaly Detection in Lidar SLAM Localization” written by Guoqing Geng et. al. presents a real-time, unsupervised anomaly detection model for Lidar Simultaneous Localization and Mapping (SLAM) stability and safety. By leveraging isolation forests and adaptive OTSU classification, the model enhances SLAM robustness in unpredictable environments. SLAM plays a critical role in smart driving systems, enabling real-time map construction and localization. However, sensor noise and dynamic environmental factors make achieving precise SLAM challenging. Errors in SLAM systems contribute to cumulative localization drift, affecting downstream tasks in autonomous driving.
The content of this manuscript aligns well with the Journal's focus and scope. The proposed method significantly outperforms state-of-the-art techniques in both Absolute Pose Error (APE) and Relative Pose Error (RPE), achieving higher trajectory accuracy and stability. It effectively reduces localization errors, mitigating the negative impact of sensor noise and external disturbances. The model’s real-time capability ensures that SLAM-based intelligent driving systems maintain reliability and accuracy in dynamic environments. This manuscript is well-composed and engaging to read. The theses presented by the Authors are supported by references to scientific articles that have been published in reputable journals. In my opinion, 39 references, most of which from the last 5 years, are a reliable point of reference for the presented results. Overall, the language used in the manuscript is clear and professional. However, there are some issues, which should be improved before its publication in WEVJ:
- Many figures are almost illegible, for example due to the font being too small. Fig. 2, Fig. 3 and Fig. 8 - 11 require improvement.
- In Section 2, there was no detailed explanation of what the correlation coefficient is. This issue may be important from the point of view of future readers of the article.
- I would suggest writing variables in italics, as this will make it easier to distinguish them from the rest of the text.
- I believe that the caption of Table 1 is too short and therefore does not provide adequate support for Readers.
- There must be a space between the number and the unit. This problem occurs, for example, in L.424.
- Why do the authors use an approximate value of Euler's constant instead of its symbol (γ)?
Author Response
Thank you for your detailed reading and effective comments. Please check the attachment for our modification.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe revised version looks better. However, the reviewer's comments have not been addressed carefully and diligently.
Author Response
Thank you for your detailed reading and effective comments. Please check the attachment for our modification.
Author Response File: Author Response.pdf
Round 3
Reviewer 1 Report
Comments and Suggestions for AuthorsThe revised paper has substantially been improved, and the authors have satisfactorily addressed the reviewer's concerns. Good attempt!