Human Fall Detection using Co-Saliency-Enhanced Deep Recurrent Convolutional Neural Networks
Chenjie Ge, Irene Yu‐Hua Gu, Jie Yang · Chalmers Research (Chalmers University of Technology) · 2019
This paper addresses issues of fall detection from videos for e-healthcare and assisted-living. Instead of using hand-crafted features from videos, we exploit a dedicated recurrent convolutional network (RCN) architecture for fall detection in combination with co-saliency enhancement. In the proposed scheme, the recurrent neural network (RNN) is realized by Long Short-Term Memory (LSTM) connecting to a set of Convolutional Neural Networks (CNNs), where each video is modelled as an ordered sequence, containing several frames. In such a way, the sequential information in video is preserved. To further enhance the performance, we propose to employ co-saliency-enhanced video frames as the inputs of RCN, where salient human activity regions are enhanced. Experimental results have shown that the proposed scheme is effective. Further, our results have shown very good test performance (accuracy 98.12%), and employing the co-saliency-enhanced RCN has led to the improvement in performance (0.70% on test) as comparing to that without co-saliency. Comparisons with two existing methods have provided further support to effectiveness of the proposed scheme.