Enhanced Isolation Forest-Based Algorithm for Unsupervised Anomaly Detection in Lidar SLAM Localization

Lidar SLAM (simultaneous localization and mapping) systems provide vehicles with high-precision maps and localization for environmental perception. However, sensor noise and dynamic changes can lead to the localization drift or localization failure of the SLAM system. Identifying such anomalies currently relies on post-trajectory analysis with subjective parameter thresholds. To address this issue, we propose an unsupervised real-time localization anomaly detection model based on the isolation forest algorithm. We first determined the necessity of variable research through variable correlation analysis. Then, we enhanced the scoring mechanism of the isolation forest by introducing a path-weighting method, improving sensitivity to complex variables and anomalies. Finally, to further increase the model’s reliability, we employed an adaptive OTSU (Otsu’s method) algorithm for automatic score classification. Experimental results show that our proposed model can effectively detect positioning anomalies by determining variable thresholds in four scenarios of the KITTI dataset. The results show excellent real-time performance, with an average running time of about 0.02 s, which is shorter than the time required to process a single data frame. Using the mean, RMSE, and standard deviation as evaluation metrics, data comparisons confirmed the algorithm’s accuracy. Compared with several SOTA (state-of-the-art) algorithms and ablation studies, our algorithm also showed higher sensitivity.