基于卷积神经网络的机器人对未知物体视觉定位控制策略

Vision-based Robot Positioning Control Strategy for Unknown Objects Using Convolutional Neural Network

无需事先获得操作目标的模型信息,仅根据当前图像特征就能实现机器人对未知目标的智能感知与定位或跟踪,一直是机器人视觉伺服领域中具有挑战性的问题.这里的"未知"不仅是指对操作台上所有物体的类别属性和位置信息未知,而且对操作物体的形状、大小等这样的几何模型先验信息也未知.为了解决这个问题,提出了一种基于卷积神经网络的机器人对未知物体的视觉定位控制策略.首先,利用基于卷积神经网络的多目标识别与检测算法获取机器人操作台上所有未知物体的类别属性信息;其次,用户根据多目标识别与检测结果在线随机地选择机器人拟定位的目标物体,此后在整个机器人视觉控制过程中,机器人视觉控制系统利用该网络从多个物体中检测出用户所选定的目标物体并计算出当前图像特征,从而实现机器人对未知目标的智能感知;最后,根据图像特征误差,设计视觉滑模定位控制律以实现机器人对未知物体的视觉定位.在MOTOMAN-SV3X六自由度工业机器人上完成了5组复杂自然场景下未知物体视觉定位实验,实验结果表明了所提出的机器人视觉定位控制策略的可行性和有效性.

The intelligent perception,positioning,and tracking of an unknown object using only the current image feature has been a challenge in the field of visual servo robot systems because of the absence of any model information about the objectprior to the task. Here, ＂ unknown＂ does not only refer to the absence of data on the classification properties and localization of all objects on the operating table,but also refers to not having the prior information on geometry such as shape and size of the manipulated objects. To solve this problem,we propose a vision-based robot positioning control strategy for unknown objects by employing the learning capability of convolutional neural network（ CNN）. First,we use a well-trained object detection network based on CNN to classify and detect objects to obtain the class labels of the objects on the operating table of the serving robot. Secondly,we select the object to be manipulated randomly by any user according to the results of automatic recognition and detection. Following this well-trained detection network,we detect the target object for subsequent images,and then compute the current image features to realize anintelligent perception for the unknown object. Finally,to achieve the robot visual positioning for unknown objects,we design a visual sliding mode positioning control law according to the image feature error to drive robotic hand/claw for expected motion. Five different experiments of robot visual positioning for unknown objects in complex natural scenes are carried out using a MOTOMAN-SV3X industrial manipulator. The experimental results confirmthe feasibility and effectiveness of the proposed robot visual positioning control strategy for unknown objects.