High-level Activity Recognition Based on Analysis of Spatio-Temporal Contexts

High-level activity recognition (HAR) has got widely used in healthcare, recommendation, activity sensitive applications etc. Because of the complexity, diversity and individuality of high-level activity, researchers are faced with the challenge of increasing recognition accuracy under realistic conditions. In daily life, high-level activity has a strong association with spatial location since people are used to doing an activity in a specific location. Furthermore, some high-level activities usually co-occur in temporal sequence, which is supposed to be considered in activity inference. In this paper, we propose an approach of high-level activity recognition based on analysis of spatio-temporal contexts. The approach extracts Region of Activity (ROA) from spatial locations and discovers Daily Activity Pattern (DAP) from temporal sequence of activities to support the design of classifiers of high-level activity. The experiments show that the proposed approach bring about promising performance improvements on activity recognition. The fusion of ROA and DAP gain an advantage over the existing work in accuracy.