LIMOT: A Tightly-Coupled System for LiDAR-Inertial Odometry and Multi-Object Tracking (2305.00406v2)

Abstract: Simultaneous localization and mapping (SLAM) is critical to the implementation of autonomous driving. Most LiDAR-inertial SLAM algorithms assume a static environment, leading to unreliable localization in dynamic environments. Moreover, the accurate tracking of moving objects is of great significance for the control and planning of autonomous vehicles. This study proposes LIMOT, a tightly-coupled multi-object tracking and LiDAR-inertial odometry system that is capable of accurately estimating the poses of both ego-vehicle and objects. We propose a trajectory-based dynamic feature filtering method, which filters out features belonging to moving objects by leveraging tracking results before scan-matching. Factor graph-based optimization is then conducted to optimize the bias of the IMU and the poses of both the ego-vehicle and surrounding objects in a sliding window. Experiments conducted on the KITTI tracking dataset and self-collected dataset show that our method achieves better pose and tracking accuracy than our previous work DL-SLOT and other baseline methods. Our open-source implementation is available at https://github.com/tiev-tongji/LIMOT.

• The paper presents a tightly-coupled system that jointly optimizes LiDAR-inertial odometry and multi-object tracking to achieve robust localization in dynamic scenarios.
• It introduces a trajectory-based dynamic feature filtering method that discards moving object features to refine scan-matching precision.
• The system employs factor graph-based optimization to adjust IMU biases and object poses, yielding improved RMSE and MOTP metrics on KITTI and custom datasets.

Overview of LIMOT: A LiDAR-Inertial Odometry and Multi-Object Tracking System

The paper "LIMOT: A Tightly-Coupled System for LiDAR-Inertial Odometry and Multi-Object Tracking" outlines the development and evaluation of a LiDAR-inertial odometry system integrated with multi-object tracking. This integration aims to address challenges in autonomous driving applications, particularly within dynamic environments where traditional SLAM algorithms based on static assumptions may falter. The proposed system, LIMOT, uses a tightly-coupled approach to jointly estimate the positions of the ego-vehicle and surrounding objects, distinguishing it from the predominant loosely-coupled methodologies in the field.

Core Contributions

LIMOT leverages a trajectory-based dynamic feature filtering method to refine scan-matching by pre-emptively identifying and discarding dynamic object features. The system is built on three primary innovations:

1. Tightly-Coupled Optimization: LIMOT integrates LiDAR-inertial odometry and multi-object tracking within a unified optimization framework, allowing simultaneous estimation of ego-vehicle and object poses. This coupling enhances localization robustness compared to methods that handle these components independently.
2. Dynamic Feature Filtering: By approximating the trajectories of tracked objects, LIMOT effectively filters out features associated with moving objects, thus enhancing the precision of scan-matching. This selective filtering distinguishes LIMOT's approach from methods which indiscriminately discard all potentially movable object features, potentially affecting data richness needed for accurate localization.
3. Factor Graph-Based Optimization: The system employs a factor graph optimization framework to further refine the biases of the Inertial Measurement Unit (IMU) and adjust the poses of both the ego-vehicle and surrounding objects over a sliding window of data. This strategy ensures consistent state estimation by efficiently handling non-static environmental elements.

Experimental Validation

The effectiveness of LIMOT is substantiated through experiments conducted on the KITTI tracking dataset and a self-collected dataset. The results indicate that LIMOT surpasses previous work (DL-SLOT and others) and related baseline methods in pose and tracking accuracy. Metrics such as the RMSE of the Absolute Trajectory Error (ATE) for both translation and rotation exhibit substantial improvements. LIMOT enhances the pose estimation of dynamic objects, supported by a higher Multi-Object Tracking Precision (MOTP) compared to conventional methods. Additionally, the performance of LIMOT in dynamic environments—like highways with numerous moving entities—underscores its suitability for autonomous driving applications in complex scenarios.

Implications and Future Directions

LIMOT demonstrates that simultaneously optimizing ego-vehicle and object poses in a SLAM context can yield notable accuracy improvements. This methodology is particularly beneficial in dynamic settings where traditional SLAM frameworks struggle. A key implication of this research lies in its potential to inform the design of autonomous navigation systems which can efficiently handle dynamic and cluttered environments.

Future research could explore the integration of dynamics models specifically tailored for moving objects, providing more granular insights into environmental interactions. There is also potential to extend this work by incorporating additional sensor modalities or testing in more diverse real-world conditions to further validate and refine the system's adaptability and performance in autonomous applications. The availability of LIMOT as an open-source project invites collaborative enhancement and cross-validation across different platforms and research contexts.