Abstract
Activity recognition (AR) methods require the ability to identify activities of crucial parts in a human body. Body segmentation techniques employed in video-based AR methods outlines significance of areas (parts) in a human body image that are essential for the follow-up methods, such as feature extraction. Existing body segmentation techniques generally make use of body modeling with or without its background and essentially entail large training datasets. These techniques cannot cope with the changes that may occur in images over time. For example, illumination and clothing changes are some of the key concerns in practical AR systems. This paper presents a new approach to employ active contour-based model that exhibits robustness to illumination and clothing changes. Existing techniques based on active contours effectively address this problem in still images and hence lack the ability to cope with illumination and changes. We present an optical flow that 1) ensures the smooth working of the model over the data that is used for estimation into contiguous frames, and 2) ensures to place the initial contour within the current frame. Our proposed approach does not depend on prior knowledge about the data and and does not require the training data, hence it is an unsupervised approach. We use the existing effective method curvelet transform to extract prominent features from activity frames. Curvelet transform keeps the line and edge information in an image intact. We employ the linear discriminant analysis (LDA) for dimension space reduction and the hidden Markov model (HMM) for activity recognition. We show that AR using our proposed method for segmentation has the accuracy comparable to those that make use of manual segmentation methods and prove its feasibility in practical AR systems.