Unmanned aerial vehicle obstacle detection and segmentation method based on dual-branch contrastive learning

By constructing a dual-branch teacher-student model for comparative learning and dynamic data augmentation, the robustness and accuracy issues of obstacle detection in UAV aerial photography under complex environments are solved, achieving high-precision, real-time obstacle segmentation, which is suitable for UAV-borne edge devices.

CN122391653APending Publication Date: 2026-07-14NANJING UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
NANJING UNIV OF SCI & TECH
Filing Date
2026-06-12
Publication Date
2026-07-14

AI Technical Summary

Technical Problem

Existing drone-based obstacle detection methods lack robustness in complex environments, have low pixel-level segmentation accuracy, and struggle to balance high precision with real-time performance requirements.

Method used

A dual-branch teacher-student model is constructed, employing both a teacher model and a student model. Through comparative learning training, combined with dynamic data augmentation and multi-level contrastive loss constraints, obstacle feature learning is enhanced. The dual-branch model is introduced only during the training phase, while the student model is retained during the inference phase, achieving high-precision pixel-level segmentation.

Benefits of technology

It significantly improves the robustness of obstacle detection and segmentation accuracy of UAVs in complex environments, while maintaining real-time performance during the inference phase and adapting to the computing power limitations of UAV onboard edge devices.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122391653A_ABST
    Figure CN122391653A_ABST
Patent Text Reader

Abstract

The application discloses a kind of based on double branch contrast learning's unmanned plane aerial obstacle detection segmentation method, belong to computer vision and unmanned plane application technical field.The application constructs the double branch instance segmentation network including teacher model and student model.Clean image and dynamic data enhancement image are respectively input into teacher and student model;By calculating the alignment loss of feature layer, detection layer and segmentation layer, consistency constraint is applied between double branches to update student model weight, and exponential moving average strategy is used to soft update teacher model weight.Inference stage discards teacher model, only retains student model to predict.The application effectively improves the feature extraction capability of model under complex illumination and occlusion conditions through dynamic data enhancement and multi-level contrast learning, realizes high-precision detection and instance segmentation of unmanned plane aerial obstacle, and does not increase the calculation overhead of actual deployment inference.
Need to check novelty before this filing date? Find Prior Art