基于神经网络修正的农业机器人视觉误差研究

Research on Visual Error of Agricultural Robot Based on Neural Network Correction

农业采摘机器人的作业是无损采摘过程,环境的复杂性与定位目标的特殊性,使得采摘机器人视觉识别过程长、误差大。传统的解决方法是分别从减少标定误差和匹配误差这两方面着手,并没有从整体的角度考虑标定和匹配的内在联系。本文提出一种视觉中的极线几何变换方法,并给出了该变换在基于BP神经网络的摄像机标定过程和基于特征的图像匹配过程中的应用分析。实验结果表明,基于极线几何变换的视觉总体误差修正方法比传统的分开修正方法具有更高的精度和实时性。

The operation of agricultural robot is a nondestructive picking process,in which the complexity of environment and particularity of location target make it very long in picking visual identity,as well as having big error. By improving calibration model and matching model respectively,traditional solution corrects those errors,without considering the internal relations of calibration and matching. This paper presents a polar geometry transform method for total visual error correction,and analyzes its application in the process of camera calibration which is based on BP neural network and image matching which is based on feature. Experimental results show that the method of total visual error correction which is based on polar geometry transform has higher precision and real-time than separated error correction in tradition.