Accidental Fall Detection Based on Pose Analysis and SVDD

摘要

With the increase of age, the risk of falling is also increasing for the elderly. The timely rescue in the event of fall can efficiently reduce the physical damage, so fall detection is a significant way to ensure the safety of the elderly. Practically, obtaining the real fall data is difficult, so the data is unbalanced. The routine fall detection models are feasible sometimes, but not universal and efficient. In order to solve this problem, we use part confident maps(PCM) and part affinity fields(PAF) to estimate the pose to get the features of joint points motion trajectory, which can well describe human motion and solve the influence of human occlusion on recognition. Then support vector data description(SVDD) is used to obtain the normal domain model of the daily behavior features to determine whether the new behavior is a fall behavior. We conduct extensive experiments on the Le2i datasets as well as a new dataset that we collect. The results demonstrate the effectiveness of the method.