中国科学院自动化研究所机构知识库

在室外开放复杂环境中，盲人出行存在很多安全隐患。利用视觉导航技术辅助盲人出行，保障盲人出行安全，对助老助残服务有重要意义，有很大的研究意义和实用价值。本文对盲人视觉辅助导航展开研究，具体研究内容如下：
（1）基于深度学习的盲道分割算法研究
针对传统视觉感知算法精度和泛化性不足的问题，本文使用基于深度学习的语义分割模型实现视觉感知，采用轻量化的语义分割模型实现盲道分割。针对数据不足问题，建立了盲道分割数据集，用于模型的训练和评估。
（2）面向语义分割的知识蒸馏算法研究
盲人辅助设备多为嵌入式移动设备，其计算资源比较有限。为了使模型适用于低算力平台，本文对面向语义分割的知识蒸馏方法展开研究，提出了关注空间多尺度差异的语义分割知识蒸馏方法，提高小模型在交通场景中的分割效果，平衡模型的准确率和速度。
（3）辅助导航中的知识表达与决策推理研究
室外出行环境开放复杂，具有高动态性和时变性，增加了导航决策的难度。
本文将出行场景中与决策有关的交通规则和生活经验抽象为知识，并将其抽象表达，用于支持视觉感知和决策。本文将盲人辅助导航抽象为智能体决策过程，基于信念-愿望-意图（BDI）模型，设计了结合知识、感知和决策的辅助导航框
架，为盲人出行提供便利。

In the open and complex outdoor environment, there are potential dangers for the blind when traveling. The use of visual navigation technology to assist the blind to travel and ensure their safety is of great significance and has great research and practical value. The main contributions are as follows:

（1） Research on Blind Track Segmentation Based on Deep Learning

This paper uses semantic segmentation based on deep learning to achieve visual perception. For the detection of blind tracks, the paper applies a lightweight semantic segmentation model. Aiming at the problem of insufficient data, a blind road segmentation data set is established for model training and evaluation.

（2）Research on Knowledge Distillation Algorithm for Semantic Segmentation

Most assistive devices for the blind are embedded mobile devices, and their computing resources are relatively limited. Considering the limited computing resources of embedded mobile devices, this paper studies the knowledge distillation for semantic segmentation, and proposes a novel semantic segmentation knowledge distillation method that focuses on spatial multi-scale differences, improves the segmentation effect of small models in traffic scenes, and balances the accuracy and efficiency.

（3）Research on Knowledge Expression and Decision Reasoning in Aided Navigation

The outdoor travel environment is highly dynamic and time-varying, which increases the difficulty of navigation decisions. The traffic rules and life experience related to decision-making in the travel scene are abstracted to knowledge, which is expressed to support visual perception and decision-making. The blind assisted navigation is abstracted into the decision-making process of the agent. Based on the belief-desire-intention (BDI) model, a framework is designed for assisted navigation combining knowledge, perception and decision-making.