Spatio-Temporal Context Kernel for Activity Recognition

Local space-time features and bag-of-feature (BOF) representation are often used for action recognition in previous approaches. For complicated human activities, however, the limitation of these approaches blows up because of the local properties of features and the lack of context. This paper addresses the problem by exploiting the spatio-temporal context information between features. We first define a spatio-temporal context, which combines the scale invariant spatio-temporal neighberhood of local features with the spatio-temporal relationships between them. Then, we introduce a spatio-temporal context kernel (STCK), which not only takes into account the local properties of features but also considers their spatial and temporal context information. STCK has a promising generalization property and can be plugged into SVMs for activities recognition. The experimental results on challenging activity datasets show that, compared to context-free model, the spatio-temporal context kernel improves the recognition performance.