Linear dynamic system method for tactile object classification

Lots of tactile sequences can be obtained by using a dexterous hand for grasping different objects. The ability of robotic environmental perception and dexterous manipulation will be significantly improved after these tactile sequences are correctly classified. Therefore, tactile sequences are separated into series of subgroups, and a method based on linear dynamical system (LDS) is used to extract features. Since these LDSs lie in non-Euclidean space, the Martin distance, which is a measurement different from Euclidean distance, is applied to calculate the distance between two LDSs, and the K-Medoid algorithm is used for clustering. The codebook is obtained after clustering and is used to represent time sequences to get a Bag-of-System (BoS). Then the BoS and labels are sent to Extreme Learning Machine (ELM) to train a classifier. Finally, three databases, KTH-7, KTH-10 and TSH-8 are used to evaluate our algorithm.