Vision-Based Fall Detection Using Dense Block With Multi-Channel Convolutional Fusion Strategy

Fall detection has become a hot issue in the field of video surveillance recently. Different from most traditional vision-based methods relying on hand-crafted features, fall detection methods based on deep learning technology can automatically mine features to detect fall events due to the powerful ability of deep learning in data analysis, and hence have received much more attention in recent years. However, information loss has become a problem that cannot be ignored, especially for the neural networks with deep layers, because loss of information will affect representativeness of features and further influence the performance of fall detection. To solve the abovementioned problem, we propose a fall detection method based on dense block with a multi-channel convolutional fusion (MCCF) strategy. In this method, MCCF-DenseBlock, a densely connected network structure, is proposed to fully extract information with its densely connected layers, and to avoid network overloading by breaking dense connections appropriately, and especially to reduce data redundancy and numerous parameters in the network via the MCCF strategy fusing its grouped features. Besides, an improved transition layer is presented to further lessen data accumulation by using a multi-level downsampling structure. Experimental results demonstrate that, the proposed method can provide accurate fall detection results (satisfactory F-score of 0.973 on the UR Fall Dataset) and outperforms several state-of-the-art methods.