Memoirs of the Scientific Sections of the Romanian Academy Tome XLI, 2018 COMPUTER SCIENCE SELECTION OF RELEVANT PARAMETERS FOR HUMAN LOCOMOTION UNSUPERVISED CLASSIFICATION SILVIU-IOAN BEJINARIU, RAMONA LUCA, FLORIN ROTARU Institute of Computer Science, Romanian Academy, Iasi Branch, Iasi, Romania Corresponding author: silviu.bejinariu@iit.academiaromana-is.ro A method for the automatic selection of the most relevant parameters for human locomotion classification is proposed. A set of 36 statistical parameters extracted from video sequences showing three basic movement types is used. Because the unsupervised classification is based on the k-means clustering algorithm, the sets of relevant parameters are determined by applying binary optimization metaheuristics using a clustering evaluation measure as objective function. Considering that the objective function is multimodal, all combinations which maximize it are retained. The binary versions of Particle Swarm and Black Hole algorithms were modified to manage the multiple solutions of the optimization process. The experiments revealed that the Black Hole algorithm leads to better results, even if it is considered a simplified version of the Particle Swarm Algorithm. Keywords: human locomotion, k-means clustering, binary optimization, nature-inspired metaheuristics 1. INTRODUCTION Human action recognition is an important task in computer vision. It has a wide spectrum of applications in many areas such as medicine, video surveillance, social activity recognition, and robotics. There are two parts in action recognition: action description, which aims at extracting motion information from video sequences, and action classification, which involves machine learning techniques to make models that assign the correct action labels. In literature there are many approaches related to the field. In [27], locomotion is classified using a neural network and data is obtained using portable devices placed on the subjects. In [1] it is proposed a medical application using pattern recognition techniques for ankle joint movement classification. Using a robust representation of spatial temporal words and an unsupervised approach during learning, [16] proposes a model to learn and recognize human actions in video. A novel method based on skeleton information provided by RGB-D cameras is proposed in [22]. Other approaches of human action recognition are proposed in [5, 19, 23, 25].