Efficient silhouette-based input methods for reliable human action recognition from videos

Abstract

In this paper, we deal the problem of classification of activities in videos by learning in a different way of modifying the usual input features to enrich the efficiency by reducing the space and time consumption. For efficient action recognition, representing important information is important for reliability. The motivation here is to extract binary silhouette (shape) information from real or raw videos. Most of the researchers are using sequence matching scheme consists of a set of successive silhouette frame that consumes much time and space for learning the system. Here, we adopted recently proposed subspace learning method: spectral regression discriminant analysis (SRDA) for dimensionality reduction. Median Hausdorff distance was used for similarity measures to match the embedded action trajectories. Then, action classification is achieved in a nearest neighbor framework. For efficient learning using SRDA, we use (i) raw silhouette, (ii) images obtained after doing distance transform (DT) in the raw silhouette, (iii) images obtained after extracting edges from the raw silhouettes—edge representation (iv) silhouette history image (SHI), and (v) silhouette energy image (SEI) that gives shape and motion information as an input feature. Using these different input methods, we achieved 100% correct recognition rate (CRR) for each cases. From the result, it is evident that SHI and SEI is found effective input method than other methods in terms of time and space consumption.