局部可视环境中基于视觉和触觉感知的灵巧手精细操作的方法研究

In order to simulate the human behavior of a target object recognition and manipulation in the local visual environment, some key scientific questions related to knowledge base construction，hybrid intelligent model construction and autonomous operation strategy of dexterous hand are proposed and solved in this project. First, to achieve the goal that the data coming from a knowledge base of the contact characteristic can be used in both stages, identification stage and operation stage, some methods of selecting and extracting contact features are proposed based on the hypothesis that the contact model of object local area is a elasticity and damping model. And the knowledge base of the contact features is constructed by using clustering algorithm; Then, a DS evidence theory and fuzzy neural network based hybrid intelligent model is constructed based on visual and tactile information fusion, to achieve target recognition of a desired object, relative pose estimation between the dexterous hand and the target object and performance evaluation of the dexterous hand operating the target object. In order to meet the validity and rapidity of contour feature extraction of a suspected target in complex environment, an algorithm of local contour reconstruction and feature extraction is proposed based on the fusion of visual and tactile information. During reconstruction, a variable step size search strategy based on active tactile perception under vision guided is adopted. At last, to improve the online manipulation performance of the object, the finger movement is planned based on a knowledge base of anthropomorphic operation, and an adaptive impedance control strategy of the contact force based on the feedback of operation performance evaluation index is proposed.

以模拟人类在局部可视环境中对目标物体进行识别和操作为背景，解决与知识库构建、混合智能模型构建和灵巧手自主操作策略相关的关键科学问题。首先，基于物体局部区域为弹性和阻尼模型假设，提出面向灵巧手操作的物体局部接触特征选定和提取方法，利用聚类算法构建接触特征知识库，实现了接触特征知识库中的数据在识别和操作两阶段复用目标；然后，构建基于D-S证据理论和模糊神经网络混合智能模型，融合视觉和触觉信息，实现期望目标物体识别、相对位姿估计和灵巧手精细操作性能评价。为解决局部可视环境中疑似目标轮廓特征提取的有效性和快速性兼顾问题，基于视觉引导主动触觉感知的可变步长搜索策略，提出了基于视觉和触觉信息融合的物体局部轮廓信息重构及特征提取算法；最后，为了实现物体操作性能的在线提升，基于拟人操作知识库规划灵巧手手指运动，提出了基于操作性能评价指标反馈的接触力自适应阻抗控制策略。

为了模拟人类在局部可视环境中对目标物体进行识别和操作，重点开展了基于触觉信息的物体重构识别及抓取点优化方法和基于知识库的触视觉融合物体识别及操作等研究。.1、提出了一种基于触觉点位置及法向量数据的未知物体模型重构方法。首先，基于视觉识别中的RBC理论，构建基于聚类、特征提取和随机森林分类的混合智能模型对物体的形状参数进行识别；其次，基于已识别的形状和触觉点，构建物体的二维形状描述符并与标准物体进行形状匹配，从而得到物体的尺寸和姿态参数；最后，通过超二次曲面方程完成对未知物体模型的构建。.2、提出了一种基于触觉探索信息的抓取点优化方法。首先，根据抓取稳定性指标从已知触觉探索数据中优选初始抓取点；然后，根据多指手运动学、可操作性约束得到目标搜索区域；最后，利用基于KNN的优化算法在目标搜索区域内搜索最优抓取点集。.3、提出了一种融合视觉和触觉信息的基于知识的物体识别新框架。该融合框架的新颖性在于，不仅将物体位姿估计紧密地集成到识别过程中，而且还提出表示物体的局部形状的最优切片概念以区别于其他物体，这可以有效提高机器人的触摸探索效率。为了模仿类人识别行为，首先在模型库和知识库的基础上，利用视觉信息找到潜在的目标，并在预识别阶段估计它们的位姿；然后在再识别阶段，规划机器人轨迹，探索每个潜在目标，提取接触点和刚度信息，计算接触刚度相似度，融合视点云和接触点云，计算最佳切片相似度；最后在决策识别阶段，确定潜在目标是否是期望物体。.4、提出了触视觉融合的拟人安全操作策略。首先，利用视觉实时检测灵巧手与期望被操作物体之间的相对位姿；然后，根据相对位姿信息，规划灵巧手运动轨迹；其次，根据已知拟人操作知识库知识，完成拟人抓取和拟人操作；最后，根据触觉传感器反馈信息，完成柔顺拟人安全操作任务。.实验或仿真结果表明，上述方法及策略的有效性。这些成果为将来类人型机器人的智能化发展提供了技术支持。