工件的视觉判别是实现机械手抓取工件的重要环节, 如何对姿态变换后的工件进行视觉识别及分类, 具有重要的理论及工程研究价值。设计了工件姿态变换实验平台, 实验研究了工件姿态变换的视觉识别问题。首先, 采用分水岭图像分割方法, 实现了工件目标图像的分割。其次, 针对工件在水平面旋转姿态变换及在三维空间进行姿态变换两种情况, 分别选取单个工件、部分遮挡工件及部分重叠工件构建图像样本, 并实现了七个特征不变矩的提取。最后, 将工件训练样本的五阶矩、七阶矩的平均值作为工件种类分类目标, 采用Parzen窗对待识别工件类型进行概率密度估计, 并用带惯性的粒子群算法对Parzen窗估计结果进行优化, 克服Parzen窗估计的多峰值问题, 实现了对工件摆放类型的准确判别。

Visual recognition of workpiece is an important part for robot to grasp the workpiece. How to recognize and classify the workpiece after attitude transformation has an important theoretical and engineering value. In this paper, the experiment platform of workpiece attitude transformation is designed, and researched the visual recognition of workpiece under pose transformation experimentally. First of all, realizes the target image segmentation of workpiece with watershed algorithm. Secondly, for the two cases as rotation pose transformation at the horizontal plane and pose transformation in three dimensional spaces, single work-piece, partial occlusion workpiece and partial overlap workpiece were selected to build image samples, and seven feature invariant moments of workpiece were extracted. Thirdly, averages of fifth-order moments and seventh-order moments of workpiece in training sample as classification target of workpiece type, the probability density of type of workpiece to be identified is estimated based on Parzen window. Used particle swarm optimization (PSO) algorithm with dynamic inertia to optimize the estimated result by Parzen window, overcome the multi peak problem of Parzen window estimation, and the accurate identification for the placed type of workpiece is realized.