The study of human activity–travel patterns for urban planning has evolved a long way in theories, methodologies, and applications. However, the scarcity of data has become a major barrier for the advancement of research in the field. Recently, the proliferation of urban sensing and location-based devices generates voluminous streams of spatio-temporal registered information. In this study, we propose an approach using the linear-chain Conditional Random Fields (CRFs) model to learn the spatio-temporal correspondences of different types of activities and the inter-dependencies among sequential activities from training dataset such as the household travel or time use surveys, and to infer the hidden activity types associated with urban sensing data. The performance of the CRFs model is compared against the Random Forest (RF) model, which has been used in a number of existing studies. The results show that the linear-chain CRFs models generally outperform the RF counterparts with respect to classification accuracy of activity types, in particular for those travelers having more outdoor daily activities. The proposed methodology is demonstrated by reconstructing the activity landscape of the surrounding area of a major Mass Rail Transit station in Singapore using the transit smart card transaction data. The inferred activities from the transit smart card data are expected to complement the ground surveys and improve our understanding of the interactions of different components of activities/travels as well as the relationship between urban space and human activities.
Error Bounds for Online Predictions of Linear-Chain Conditional Random Fields: Application to Activity Recognition for Users of Rolling Walkers
