Fuzzy Image Matching for Posture Recognition in Ballet Dance Sriparna Saha †1 , Anupam Banerjee ¥2 , Sumana Basu ¥3 , Amit Konar †4 † Electronics & Telecommunication Engineering Dept. ¥ Computer Science and Engineering Dept. Jadavpur University, India { 1 sahasriparna, 2 mr.anupambanerjee, 3 sumana.basu21}@gmail.com, 4 konaramit@yahoo.co.in Atulya K. Nagar Mathematics and Computer Science Dept. Liverpool Hope University United Kingdom nagara@hope.ac.uk Abstract— This work aims at designing a fuzzy matching algorithm that would automatically recognize an unknown ballet posture from seventeen fundamental ballet dance primitives. A novel and simple 7-stage system is proposed to achieve the desired objective. Minimized skeletons of the dance postures are generated after performing skin color segmentation on them. Straight line approximation on the minimized skeletons with the help of chain code and sampling generate their equivalent stick figure diagrams. Significant straight lines from the stick figure diagrams are considered, their fuzzy membership with respect to the 4 quadrants are evaluated. Finally with the help of the evaluated data, a fuzzy T-norm operator determines the proximity of a generated dance posture with the seventeen fundamental dance primitives. Keywords— chain code; fuzzy T-norm; membership function; skeleton; straight line approximation I. INTRODUCTION Ballet is a technical form of dance, well known for its elegant and graceful movements. We can trace back its roots to Italy, where this dance form originated in the fifteenth century. Cultural transfusion, especially from countries like France and Russia has influenced this dance form to a considerable extent. Ballet consists of 17 static postures, namely arabesque, arms first, arms second, arms third, arms fourth, arms fifth, attitude front, attitude back, attitude side, croise derriere, croise devant, ecarte devant, efface devant, en face, posture front, posture back, and releve. The proposed algorithm focuses on fuzzy image matching for posture recognition of ballet dance, thereby facilitating e- learning of the said dance procedure. E-learning comes with its own set of advantages, the primary ones being the ease of learning and the flexibility it offers to its customers. The cost of learning decreases considerably and the active interaction between the software and the dancer makes the dance form even more intriguing. In [1], the authors sequence the ballet dance moves using stick figure diagrams. Here human motion sensor device is used, as a result of which the technique becomes expensive and unfit for e-learning. Another algorithm proposes placing of two cameras orthogonally at a mid-body height for dance gesture recognition [2]. The cumbersome arrangement makes this technique unfit for e-learning purposes. Additionally, owing to its generality, [2] lacks the necessary acumen to deal with specific dance forms like ballet. In [3], dance datasets are prepared using single and multiple gestures and they are classified using hidden Markov model. For 3D projection of dance gestures, multiple calibrated cameras are needed, which are costly and hence not suitable for the purpose of e-learning. In this work a single camera suffices the entire procedure. Markov model along with K-means algorithm is used for removal of ambiguity in posture recognition in [4]. For posture recognition, the background in which the performance takes place plays an important role. Noisy background, improper texture of the dress and an inappropriate distance of the dancer from the camera leads to erroneous results. The authors in [5] propose an algorithm using histograms to successfully deal with the above mentioned problems. In [6], the authors detect basic human postures with the help of active contours and neural networks. Here background subtraction technique is implemented at the pre processing stage. But the effectiveness of the algorithm is limited to only three postures- sitting, bending and squatting. Another approach used for human posture recognition uses decision tree and is discussed at length in [7]. Decision tree is based on entropy which measures information gain [8]. A fuzzy theory based approach to posture recognition uses a genetic fuzzy finite state machine. Here inputs to the state machine are provided and outputs are obtained using fuzzy inference rules [9].Comparison of supervised and unsupervised learning classifiers for human posture recognition is undertaken in [10]. For the identification of ballet postures, the first problem that arises is segmentation of the dancer from the background. The dance is performed adhering to a specific dress code, by virtue of which the head, hands and legs of the dancer are distinctly visible. For this reason skin color based segmentation is applied. The original images of dance postures are in RGB color space and as a result are susceptible to the problems of varying illumination. To make the images illumination invariant, the images are converted to HSV and YCbCr color space. By assigning limits to Cb and Cr values, hue value