融合RGB特征和Depth特征的3D目标识别方法

3D object recognition method based on the fusion of RGB and depth feature

针对目标类内差异、类间相似的识别问题,结合RGB图像和Depth图像各自的优势,提出一种基于多核学习的融合RGB特征和Depth特征的3D目标识别方法。该方法提取目标物体的RGB特征和Depth特征;并根据两种特征的类内、类间相似性均值和方差,为特征自适应的分配不同的权重;最后利用多核学习(MKL)的方法对特征进行加权融合,并结合SVM分类器,实现3D目标识别。最后通过在Kinect相机得到的RGB-D数据集上进行实验,验证了该文方法能够有效地实现对RGB特征和Depth特征的融合,很好的解决类内差异、类间相似的3D目标识别问题,提高了3D目标识别的识别率。

For solving the differences of intra-class and inter-class similarity problems,combining with the respective advantages of RGB image and Depth image,a new 3D object recognition method based on the fusion of RGB and Depth feature using Multiple Kernel Learning（MKL） is proposed by extracted RGB and Depth Feature from3 D object.Then,based on the similarity mean and variance of intra-class and inter-class,the different weight is assigned adaptively according to the contribution degree of recognition result of two features.Finally,by using MKL to achieve a weighted feature fusion and realize the recognition of 3D object with Support Vector Machine（SVM）classifier.Finally,the experiments are conducted on the RGB-D data set collected by Kinect camera.It validates that the proposed method can effectively combine RGB and Depth features,solve the differences of intra-class and inter-class similarity problems on 3D object recognition,and improve the 3D object recognition rate.