基于增量式双向主成分分析的机器人感知学习方法研究

针对直观协方差无关增量式主成分分析算法(CCIPCA)需要满足零均值高斯分布的问题，该文提出含均值差向量更新的泛化CCIPCA算法(GCCIPCA)，拓展了算法的适用范围。其次，针对机器人感知学习存在的在线增量计算及有效数据降维等问题，将GCCIPCA的增量思想引入到现有的双向主成分分析算法(BDPCA)，提出基于增量式BDPCA(IBDPCA)的机器人感知学习方法。该方法直接针对图像矩阵行列方向的类散度矩阵进行迭代估计，具有一定的泛化能力和快速的增量学习能力，提高了实时处理速度。最后，以机器人待抓取物块作为感知对象进行实验，结果表明所提算法能够满足机器人感知学习的实时处理需求，相比现有的增量式主成分分析算法，在收敛率、分类识别率、计算时间及所需内存等性能方面均得到显著提升。

Robot Perceptual Learning Method Based on Incremental Bidirectional Principal Component Analysis

Existing Candid Covariance-free Incremental PCA (CCIPCA) has the limitation of the stable image inherent covariance, and a Generalized CCIPCA (GCCIPCA) with an appended term of the mean difference vector is presented. It can be considered that the CCIPCA is only a special case of the GCCIPCA and can extend the scope of the algorithm. Then, the incremental learning of the proposed GCCIPCA is innovated to the existing Bi-Directional PCA (BDPCA), and the called Incremental BDPCA (IBDPCA) is used for the robot perceptual learning and it can be used to incrementally compute the principal components without estimating the similar scatter matrixes in the row and column directions, which can build up the real-time processing speed greatly. Finally, the blocks grasped by the robot are used as the perceptual objects, and the experimental results demonstrate that the proposed algorithm works well, and the convergence rate, the classification recognition rate, the computation time and the required memory are improved significantly.