InsMOS: Instance-Aware Moving Object Segmentation in LiDAR Data (2303.03909v1)

Abstract: Identifying moving objects is a crucial capability for autonomous navigation, consistent map generation, and future trajectory prediction of objects. In this paper, we propose a novel network that addresses the challenge of segmenting moving objects in 3D LiDAR scans. Our approach not only predicts point-wise moving labels but also detects instance information of main traffic participants. Such a design helps determine which instances are actually moving and which ones are temporarily static in the current scene. Our method exploits a sequence of point clouds as input and quantifies them into 4D voxels. We use 4D sparse convolutions to extract motion features from the 4D voxels and inject them into the current scan. Then, we extract spatio-temporal features from the current scan for instance detection and feature fusion. Finally, we design an upsample fusion module to output point-wise labels by fusing the spatio-temporal features and predicted instance information. We evaluated our approach on the LiDAR-MOS benchmark based on SemanticKITTI and achieved better moving object segmentation performance compared to state-of-the-art methods, demonstrating the effectiveness of our approach in integrating instance information for moving object segmentation. Furthermore, our method shows superior performance on the Apollo dataset with a pre-trained model on SemanticKITTI, indicating that our method generalizes well in different scenes.The code and pre-trained models of our method will be released at https://github.com/nubot-nudt/InsMOS.

Instance-Aware Moving Object Segmentation in LiDAR Data

The paper introduces "InsMOS: Instance-Aware Moving Object Segmentation in LiDAR Data," presenting an innovative approach for segmenting moving objects in 3D LiDAR scans, crucial for autonomous navigation systems. The objective is to enhance how autonomous systems perceive dynamic environments by distinguishing between moving and static objects, a fundamental challenge for applications such as SLAM, dynamic map updating, and real-time path planning.

Technical Approach and Contributions

The proposed method incorporates a novel network architecture that not only assigns point-wise moving labels in a LiDAR scan but also identifies instance information of key traffic participants. This dual focus on motion and instance segmentation facilitates a more nuanced interpretation of a scene, enabling differentiation between truly moving objects and those that are temporarily stationary.

Key contributions include:

• 4D Sparse Voxel Processing: The method processes sequences of point clouds as 4D voxels, leveraging 4D sparse convolutions to extract motion features and integrate them with the current LiDAR scan.
• Spatio-Temporal Feature Extraction: From the updated scan, spatio-temporal features are derived for the purpose of instance detection and subsequent feature fusion.
• Upsample Fusion Module: Designed to integrate spatio-temporal features with predicted instance data, this module outputs fine-grained point-wise segmentation to accurately label moving and static objects.
• Benchmark Evaluation and Performance: On the LiDAR-MOS benchmark using SemanticKITTI, the approach surpasses state-of-the-art methods in moving object segmentation. Furthermore, it demonstrates robust performance on the Apollo dataset with only a pre-trained SemanticKITTI model, illustrating adaptability across different driving environments.

Numerical Results and Implications

Achieving superior IoU scores on the LiDAR-MOS benchmark indicates this method's effectiveness in identifying motion contexts, a critical capability for autonomous agents navigating complex and dynamic urban spaces. The method's successful generalization on different datasets without extensive fine-tuning signifies its potential adaptability and broad application scope.

Theoretical and Practical Implications

Theoretically, the incorporation of instance information in motion segmentation marks an advancement in bridging object-level understanding with point-based processing in LiDAR data. Practically, this research has the potential to greatly enhance autonomous driving capabilities by enabling real-time decision-making based on accurate and detailed environmental perception.

Future Directions

Looking forward, further integration with multiple sensor modalities, such as cameras and radars, could heighten robustness and accuracy. Additionally, extending this approach could involve augmented reality applications where differentiating between static and moving objects seamlessly aligns virtual and real-world environments.

In summary, "InsMOS" delivers significant strides in LiDAR-based segmentation, integrating motion analysis with instance recognition to push the boundaries of autonomous navigation and perception systems.