SMDFusion: Multimodal Sensor Fusion for Auto Driving Perception Under Mixed-Traffic Scenarios

View at publisher Summarise with VIRAT
Authors: Ting Qu, Dewang Liu, Jinghua Zhao, Hong Chen
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Autonomous driving relies on accurate and reliable perception particularly in mixed traffic scenarios. In such complex environments including vehicles, pedestrians, barriers, and significant occlusions occur, degrading the quality of information collected by LiDAR and cameras. Therefore, current multimodal sensor fusion frameworks face difficulties in capturing long-range dependencies in point clouds and integrating information from sensors of different quality, which leads to inaccurate localization and object misclassification. To address these issues, we propose SMDFusion, a novel multimodal framework that effectively fuses LiDAR features, image frequency features, and depth information. Specifically, a Sparse-Mamba (SMM) module is introduced to capture long-range dependencies in sparse voxel features via Hilbert curve serialization and state-space modeling; a Depth-guided Frequency Enhancement (DFE) module leverages wavelet decomposition to refine image feature representations; and a Dynamic Gated Multimodal Fusion (DGMF) module adaptively models spatial, channel, and modality interactions for robust fusion. In addition, a distribution classification loss is designed to optimize the decision boundaries and mitigate misclassification among visually similar categories. Experiments on the large-scale nuScenes benchmark demonstrate that SMDFusion achieves 74.6% NDS and 71.9% mAP on the test set without test-time augmentation or model ensemble, while ablation studies further validate the contribution of each component.

Fields of Study & Subjects

Autonomous Vehicle Technology and Safety Advanced Neural Network Applications Advanced Vision and Imaging

Publication details

Journal / source
IEEE Transactions on Intelligent Transportation Systems
Pages
1–14
Publication date
1 January 2026
Language
EN