Thermal imaging based elderly fall detection

Elderly fall detection is very special case of human action recognition from videos and has very practical application in old age home and nursing centers. Fall detection in its simplest form is a binary classification of fall event or other daily routine activities. Hence, the current trend of sophisticated techniques being developed for human action recognition, particularly with scenarios of large number of classes may not be required in elderly fall detection. However, other design considerations such as simplicity (ready to be deployed), privacy issues (not revealing the identity) are to focused and are the major contributions of this paper. The Spatio-Temporal Interest Points (STIP) and Fisher vector framework for human action recognition is established as baseline in this work. A novel optical flow based technique is proposed that yields better performance than the baseline. Further, a very economical thermal imaging based input modality is proposed. Along with the thermal images not revealing the identity of the persons, thermal images also aid human detection from backgrounds – a useful solution in computing the optical flow of human movements. The proposed solution is also validated on the KUL Simulated Fall dataset showing its generalization capability.