Surface Textures Classification with Fractal Detrended Fluctuation Analysis

Tactile perception provides robots and prosthetics with capabilities such as object recognition, precise manipulation, and natural interaction. Tactile feedback plays an important role by continuously providing individuals with vital information about their external environment through physical contact. Therefore, rapid developments in human-friendly biomimetic electronics and flexible devices enable robots to distinguish material properties such as local geometry and texture, especially for materials such as textiles. In this paper, a new method for surface texture classification based on tactile signals is proposed. In the proposed method, firstly, 3-axis accelerometer (X, Y, Z) tactile signals and microphone signals are subjected to data augmentation with a non-overlapping sliding window approach. Feature extraction is performed with Fractal Detrended Fluctuation Analysis (FDFA), which is an effective method for investigating long-term correlations of power law of non-stationary time series. In the last stage, the textures were classified by using the Support Vector Machine (SVM), a widely preferred machine learning algorithm, using features obtained from accelerometer and microphone signals separately and combined. Experimental results show that when the window length is selected as 1 second, 82.91% classification accuracy is achieved for accelerometer data, 98.33% for microphone data, and 99.16% for the combined use of data from both sensors. Compared to studies in literature, 12.08% higher classification performance is achieved for microphone data and 0.56% higher classification performance is achieved when accelerometer-microphone data are combined.