Discriminative sparse projections for activity-based person recognition

In this paper, we propose an activity-based person recognition approach based on discriminative sparse projections (DSPs) and ensemble metric learning. Unlike gait recognition where only the walking activity is utilized for human identification, we aim to recognize people from more types of activities such as eating, drinking, running, and so on. Our motivation is inspired by the fact that people do not always walk in person identification systems and gait recognition could fail to work in this scenario. Given each video clip, we first extract the binary human body mask in each frame by using background substraction. Then, we propose a DSP method to project these body masks into a low-dimensional subspace and cluster them into a number of clusters simultaneously. Subsequently, each video clip is pooled as a histogram feature for activity representation. Lastly, we propose an orthogonal ensemble metric learning (OEML) method to learn a distance metric to exploit more discriminative information for recognition. Experimental results on five benchmark activity databases are presented to show the efficacy of our proposed approach. Moreover, we apply our approach to gait recognition and also achieve competitive results with the state-of-the-art methods on the widely used CASIA-B gait dataset.