Transformer-Based Optimized Multimodal Fusion for 3D Object Detection in Autonomous Driving

Abstract

Accurate 3D object detection is vital for autonomous driving since it facilitates accurate perception of the environment through multiple sensors. Although cameras can capture detailed color and texture features, they have limitations regarding depth information. Additionally, they can struggle under adverse weather or lighting conditions. In contrast, LiDAR sensors offer robust depth information but lack the visual detail for precise object classification. This work presents a multimodal fusion model that improves 3D object detection by combining the benefits of LiDAR and camera sensors to address these challenges. This model processes camera images and LiDAR point cloud data into a voxel-based representation, further refined by encoder networks to enhance spatial interaction and reduce semantic ambiguity. The proposed multiresolution attention module and integration of discrete wavelet transform and inverse discrete wavelet transform to the image backbone improve the feature extraction capability. This approach enhances the fusion of LiDAR depth information with the camera’s textural and color detail. The model also incorporates a transformer decoder network with self-attention and cross-attention mechanisms, fostering robust and accurate detection through global interaction between identified objects and encoder features. Furthermore, the proposed network is refined with advanced optimization techniques, including pruning and Quantization-Aware Training (QAT), to maintain a competitive performance while significantly decreasing the need for memory and computational resources. Performance evaluations on the nuScenes dataset show that the optimized model architecture offers competitive results and significantly improves operational efficiency and effectiveness in multimodal fusion 3D object detection.

Keywords

• Autonomous driving
• LiDAR
• multimodal fusion
• network compression
• pruning
• quantization