Human Gait Analysis and Recognition Using Support Vector Machines

Human gait reveals feelings, intentions and identity which is perceived by most human beings. To understand this perceptual ability, Swedish psychologist Gunnar Johansson (1973), devised a technique known as PL (Point Light) animation of biological motion. In his work, the activity of a human is portrayed by the relative motions of a small number of markers positioned on the head and the joints of the body. This paper explores the basic concept of PL animation along with machine vision and machine learning techniques to analyze and classify gait patterns. Basically, frames of each video are background subtracted, the silhouette noise found were salt noise and noise connected in large blobs which are detected and removed based on morphological operations and area of connected components respectively. Image is then segmented and body points such as hand, knee, foot, neck, head, waist along with the speed, height, width, area of person are determined by an algorithm. We then fit sticks connecting pairs of points. The magnitude and direction of these stick along with other features forming a 24-dimensional feature vector for each frame of a video are classified using SVM Modelling using LIBSVM Toolkit. The maximum recognition accuracy found during testing by cross validation with parameters of LIBSVM was 93.5 %.