基于多模态融合的RGB-D物体识别

RGB-D Object recognition based on Multimodal Fusion

针对现有的RGB-D物体识别方法存在特征学习不全面导致物体识别准确率不高的问题,结合分层匹配追踪算法(Hierarchical matching pursuit,HMP)和特征级融合提出一种改进的物体识别算法。该算法首先利用稀疏编码和池化技术分别从RGB-D图像(RGB图像和深度图像两种模态)中提取RGB特征和深度特征,然后根据不同模态的特征对物体识别率的贡献进行特征级融合得到多模态融合特征,最后送入SVM分类器进行分类识别,并调整融合参数寻求最优识别率。在RGB-D数据集上进行分类识别实验,结果表明该方法的物体分类识别率能够达到83.6%,比其他方法提高了1%-2%。

For the problem that insufficient feature learning lead to lower accuracy of object recognition for the existing RGB-D object recognition methods, an improved object recognition method base on a combination of hierarchical pursuit pursuit algorithm(HMP) and feature level fusion was proposed. The algorithm firstly combines sparse coding with pooling techniques to extract distinctive RGB feature and depth feature from RGB-D images(two modes of RGB image and depth image), and then adopt feature level fusion method to obtain multimodal fusion feature according to the contribution of different modal feature to object recognition rate. Finally, send the multimodal fusion feature to SVM classifier for classification recognition, and adjust the fusion parameters for the best recognition rate. The classification recognition experiment on RGB-D dataset shows that the object recognition rate of this method can reach 83.6%, which is 1%-2% higher than other methods.