Real-time general object recognition for indoor robot based on PCL

We present a system for real-time general object recognition (gor) for indoor robot in complex scenes. A point cloud image containing the object to be recognized from a Kinect sensor, for general object at will, must be extracted a point cloud model of the object with the Cluster Extraction method, and then we can compute the global features of the object model, making up the model database after processing many frame images. Here the global feature we used is Clustered Viewpoint Feature Histogram (CVFH) feature from Point Cloud Library (PCL). For real-time gor we must preprocess all the point cloud images streamed from the Kinect into clusters based on a clustering threshold and the min-max cluster sizes related to the size of the model, for reducing the amount of the clusters and improving the processing speed, and also compute the CVFH features of the clusters. For every cluster of a frame image, we search the several nearer features from the model database with the KNN method in the feature space, and we just consider the nearest model. If the strings of the model name contain the strings of the object to be recognized, it can be considered that we have recognized the general object; otherwise, we compute another cluster again and perform the above steps. The experiments showed that we had achieved the real-time recognition, and ensured the speed and accuracy for the gor.