Clustering and recognition for automated tracking and grasping of moving objects

Catching moving and random placed packages of different styles on conveyer belt by a robot is important and worthy of study for improving efficiency. The characters, color and the texture etc. on their envelope become very useful information to classify and recognize them. In this paper, we extract local features (SIFT) from the image sequences of packages' visual appearance, and map these keypoints of each image into a unified dimensional histogram vector (Bag-of-words) after clustering, this histogram is treated as an input vector for a multi-class SVM to build the training classifier model and recognize the moving object real time. In operation of tracking a certain target, it is necessary to acquire its position and orientation by matching SIFT features of adjacent images, and obtain gripping region of robot by estimating the object's position and orientation using Kalman filter. For organizing the wide variety of high-dimension local features automatically and acquiring accurate classification, we proposed a modified fuzzy c-means algorithm based on the particle swarm optimization and shadowed sets (SP-FCM) to perform feature clustering. Experimental results show that the SP-FCM algorithm is effective and the moving packages are classified, tracked, and stably grasped.