A patrol robot based on multi-modal data fusion analysis
CN121170741BActive Publication Date: 2026-08-28广州南沙海港集装箱码头有限公司 +1
Patent Information
- Application Number
- CN202511555476.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-29
- Publication Date
- 2026-08-28
- Estimated Expiration
- 2045-10-29
AI Technical Summary
Technical Problem
[0004]本发明解决的技术问题是：现有技术易受低照度、强光、阴影、遮挡、灰尘、烟雾等环境因素影响，导致检测盲区或误判
Benefits of technology
[0015]本发明的有益效果：本申请结合了反色处理、基于m叉树和遗传算法优化的全局大气光估计，以及自适应窗口的暗通道先验结合引导滤波计算透射率，实现了在复杂光照（特别是低照度）环境下高质量、高鲁棒性的图像恢复，保证了视觉输入的有效性。引入短时能量、频谱质心和频谱扩散等多维声学特征，并提出动态门槛设定和基于背景噪音评估的异常事件起止点判断机制，这使得巡检机器人能够智能适应不同环境噪音，精准地从嘈杂环境中分离出真正有意义的异常声音事件，提高了声学分析的准确性和抗干扰能力；将分数阶傅里叶变换与递归图编码应用于一维声音信号，将其高效、有意义地转化为二维图像特征图，这使得异构的声学数据能够与视觉数据（可见光、红外）在同一维度和特征空间中进行深度融合，克服了传统方法中声学与视觉数据融合的维度和语义鸿沟，为后续的深度学习模型处理提供了统一的输入。本申请通过计算各模态特征图的梯度作为权重，进行加权融合，这种策略确保了融合后的特征图能够最大程度地保留所有模态中最清晰、最显著的边缘、纹理和模式边界信息。这种精细化的融合方式使得融合特征更具判别力，显著提升了后续识别和定位的精度和鲁棒性。
✦ Generated by Eureka AI based on patent content.
Smart Images
Figure CN121170741B_ABST
Abstract
The application discloses a kind of based on multimodal data fusion analysis's inspection robot, it is related to the technical field of image communication, data pre-processing module is used to carry out enhancement processing to the visible light image and infrared image of acquisition, obtains enhanced image, and the sound signal of acquisition is analyzed, and the sound segment including abnormal sound event is segmented out;Feature fusion module is used to convert sound segment into two-dimensional acoustic feature map, extracts the deep feature map of enhanced image, according to the gradient information of two-dimensional acoustic feature map and deep feature map, the features of different modalities are weighted and fused, and generate multimodal fusion feature map;Analysis decision module is used to generate inspection report.This application constructs a fusion visual and auditory inspection robot intelligent analysis system, and through a series of data preprocessing, cross-modal feature conversion and multi-level feature fusion method, aims at overcoming the limitations of single mode in complex inspection environment.
Need to check novelty before this filing date? Find Prior Art
Citation Information
Patent Citations
Inspection robot and inspection method
CN113538723A
Speech enhancement method based on multi-domain fusion network low-cost automatic search construction
CN117037831A
Low-illumination image enhancement method for simulating foggy day degradation mechanism
CN119850493A
Transformer fault intelligent detection method and device based on deep learning multi-mode fusion, computer equipment and readable storage medium
CN120632670A