Multiple Views Based Recognition of Human Activities using Uniform Patterns

Great efforts are made for the recognition of a person’s activity, still it is challenging research domain in security and surveillance. This paper proposes an efficient framework to recognize activities captured from multiple views by incorporating cameras placed at different viewing angle. These viewpoints may be horizontal, vertical, top-down as well as several others. Sometimes three cameras are placed for this purpose whereas sometimes number of cameras may be five or eight. The framework includes 3 consecutive modules that are: to locate humans in a video using background subtraction method, to extract uniform LBP and to classify human actions/activities using SVM multiclass classifier with OVA architecture. The rotation invariant characteristic of LBP supports in human activity classification from multiple views. In addition to this, better discrimination capability of these patterns provides high efficiency to the proposed framework. A hierarchical classification technique has been implemented and multiple SVMs are aggregated to classify human activities. Experimentation was performed on CASIA and IXMAS activity datasets and demonstrates the effectiveness of the proposed framework for multiple views.