Multimodal 3D target detection method based on knowledge distillation category adaptive fusion

By designing a category-aware adaptive fusion network and a knowledge distillation framework, the problems of poor fusion and high computational complexity in multimodal 3D object detection are solved, achieving high-precision and efficient target detection for autonomous driving.

CN122244835APending Publication Date: 2026-06-19TAIYUAN UNIVERSITY OF SCIENCE AND TECHNOLOGY
0 Cites 0 Cited by

Patent Information

Application Number
CN202610407950.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-03-31
Publication Date
2026-06-19

AI Technical Summary

Technical Problem

Existing point cloud-based 3D object detection frameworks suffer from problems such as poor multimodal information fusion, high computational complexity, long training time, and difficulty in model deployment in autonomous driving, and thus cannot meet practical needs.

Method used

We design a category-aware adaptive fusion network, which adaptively adjusts the grid by using density-aware angle partitioning and distance-aware radial partitioning weight prediction modules, and combines them with a point weight prediction module to achieve feature aggregation. We construct a teacher-student knowledge distillation framework for lightweight processing, and use logits distillation, feature distillation and label distillation for multimodal feature fusion.

Benefits of technology

It improves the accuracy and robustness of multimodal 3D target detection, achieves a balance between detection efficiency and performance, and reduces model complexity and computational overhead.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention provides a multi-modal 3D target detection method based on knowledge distillation category adaptive fusion, and belongs to the technical field of 3D target detection methods, and the method comprises the following specific steps: S1, constructing a multi-modal 3D target detection network CAAF-DET3D based on category perception adaptive fusion, carrying out multi-modal feature fusion on different detection categories at a cylindrical coordinate view angle by adopting a self-adaptive grid and a self-adaptive weight; s2, training on a public large data set by using a CAAF-DET3D network to obtain a teacher model; s3, constructing a knowledge distillation framework of the teacher model and the student model, and guiding the student model to train by adopting logits distillation, feature distillation and label distillation; and S4, performing target detection by using the trained student model. According to the method, the reasoning efficiency is remarkably improved on the premise of ensuring the detection precision, and the efficiency and the performance of the multi-modal 3D point cloud target detection method are balanced.
Need to check novelty before this filing date? Find Prior Art