Indoor Activity Recognition Using a Hybrid Generative-Discriminative Approach with Hidden Markov Models and Support Vector Machines

Human activity recognition is used for many practical applications such as context modeling in smart cities, surveillance and assisted living. In this paper, we apply a hybrid generative-discriminative approach using Fisher kernels with inverted Dirichlet-based and inverted Beta-Liouville-based hidden Markov models (HMMs) to improve the recognition performance. We propose a method that combines HMMs as a generative approach, with the discriminative approach of Support Vector Machine (SVM). This strategy allows us to deal with Spatio-temporal motion data, and at the same time use the special focus on the classification task that SVM could provide us. Experiments on the challenging activity recognition benchmark UCF101, demonstrate an effective improvement of the recognition performance compared to the standard generative and Gaussian-based HMM approaches.

I. Introduction

Human activity recognition has recently gained prominent interest as an active area of computer vision research. Indeed, being able to accurately identify different performed activities allows for further understanding of people’s lifestyle and their needs over a period of time. Ranging from assisted living [1], [2], elderly fall detection [3], to smart homes technologies [4], applications of activity recognition (AR) are numerous and equally critical.