Human Activity Recognition using Max-Min Skeleton-based Features and Key Poses

Abstract

Human activity recognition is still a very challenging research area, due to the inherently complex temporal and spatial patterns that characterize most human activities. This paper proposes a human activity recognition framework based on random forests, where each activity is classified requiring few training examples (i.e. no frame-by-frame activity classification). In a first approach, a simple mechanism that divides each action sequence into a fixed-size window is employed, where max-min skeleton-based features are extracted. In the second approach, each window is delimited by a pair of automatically detected key poses, where static and max-min dynamic features are extracted, based on the determined activity example. Both approaches are evaluated using the Cornell Activity Dataset [1], obtaining relevant overall average results, considering that these approaches are fast to train and require just a few training examples. These characteristics suggest that the proposed framework can beuseful for real-time applications, where the activities are typicallywell distinctive and little training time is required, or to be integrated in larger and sophisticated systems, for a first quick impression/learning of certain activities