MonoAMP: Adaptive Multi-Order Perceptual Aggregation for Monocular 3D Vehicle Detection

Area of Science:

• Computer Vision
• Autonomous Driving Systems

Background:

• Monocular 3D object detection is crucial for autonomous driving due to efficiency.
• Existing methods struggle with cross-dimensional attention and multi-order context in complex scenes.

Purpose of the Study:

• To introduce MonoAMP, an adaptive multi-order perceptual aggregation algorithm.
• To enhance feature attention and contextual information modeling for improved 3D object detection.

Main Methods:

• Implemented triplet attention for enhanced cross-dimensional feature interaction.
• Developed an adaptive multi-order perceptual aggregation module for dynamic context capture.
• Proposed an uncertainty-guided adaptive depth ensemble for robust depth prediction.

Main Results:

• MonoAMP achieved significant performance gains on the KITTI dataset (16.80% AP3D, 24.47% APBEV).
• Ablation studies showed a 3.78% improvement in object detection accuracy over baseline methods.
• Demonstrated superior detection capabilities, particularly in complex driving scenarios.

Conclusions:

• MonoAMP effectively addresses limitations in current monocular 3D object detection methods.
• The proposed adaptive aggregation and uncertainty-guided depth fusion enhance perception and accuracy.
• MonoAMP offers a promising solution for robust 3D object detection in autonomous driving.