Appearance and Depth for Rapid Human Activity Recognition in Real Applications

Human activity recognition has gained a lot of attention in the computer vision society, due to its usefulness in numerous contexts. This work focuses on the recognition of Activities of Daily Living (ADL), which involves recordings constrained to specific daily activities that are of interest in assisted living or smart home environments. We present a novel technique for spatial activity localisation and recognition from colourdepth sequences, tailored to Activities of Daily Living (ADLs), which usually take place in relatively constrained environments. The proposed method significantly reduces the computational cost of activity recognition, while at the same time achieving a competitive accuracy rate, comparable to the State of the Art (SoA). This is achieved by the introduction of appearance and depth based spatiotemporal volumes, the Spatio-Temporal Activity Cells (STACs), extracted using appearance and depth information from successive video frames. A novel adaptive background modelling method follows, to characterize the STACs as “active” or “inactive” and accumulate them into foreground or background history volumes respectively. After activity detection using the STACs, activity recognition takes place using a novel, depth-based descriptor, the Histogram of Surface Normals Projections (HONSP), in combination with well known appearance descriptors (Histograms of Oriented Gradients, HOGs). Fisher encoding aggregates them into a fixed size vector to train a multiclass SVM model, which is then used for activity recognition. Experiments on different ADL datasets recorded with elderly people verify that the suggested algorithm is very appropriate for real life scenarios.