BEV-Nexus: BEV Perception Algorithm Based on Depth Perception Enhancement and Dynamic Adaptive Fusion

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Existing 3D object detection methods struggle with accurate depth estimation and efficient fusion of image and point cloud data.
• Alignment deviations in Bird's-Eye View (BEV) features hinder the performance of multimodal fusion frameworks.

Purpose of the Study:

• To propose an improved multimodal fusion framework, BEV-Nexus, for accurate and efficient 3D object detection.
• To address limitations in depth prediction and feature fusion within image-point cloud fusion methods.

Main Methods:

• Introduced a Point-Cloud-Guided Depth Prediction Network (PCGD-Net) for enhanced image-based depth estimation.
• Designed a Dynamic Self-adaptive Feature Fusion Module (DSF-Module) using window attention for robust multimodal feature fusion.
• Incorporated a Dilated Attention Enhancement Block (DAEB) with dilated convolution and SimAM for improved feature representation.

Main Results:

• BEV-Nexus achieved 1.8% higher mAP and 1.5% higher NDS compared to BEVFusion on the nuScenes validation set.
• On the nuScenes test set, BEV-Nexus improved mAP by 1.6% and NDS by 1.4%.
• The framework demonstrated significant lightweight advantages with detection FPS remaining nearly unchanged.

Conclusions:

• BEV-Nexus offers a substantial improvement in 3D object detection accuracy and efficiency.
• The proposed PCGD-Net and DSF-Module effectively address depth estimation and feature fusion challenges.
• The framework provides a lightweight yet high-performance solution for real-world autonomous driving applications.