Daily activity prediction based on spatial-temporal matrix for ongoing videos

Predicting human activities is an emerging area in computer vision to help computational systems to forecast ongoing human activities. This paper recognizes the challenging activities of daily living that all contain similar manipulations. A spatial-temporal matrix (STM) feature descriptor is used to involve the shape and motion information. Then, a temporal bag-of-words algorithm is proposed to interpret the local feature vectors on the diagonal of the STM and support vector machine (SVM) is performed to train the classifier. Experimental results show that the proposed approach outperforms the state-of-the-art activity prediction algorithms. It also proves this framework can intuitively represent the image sequences of ongoing human activities. This can be a benefit to the applications on human-machine interaction (HMI).