Real-time multimodal ADL recognition using convolution neural networks

摘要

Activities of daily living (ADLs) are the activities which humans perform every day of their lives. Walking, sleeping, eating, drinking and sleeping are examples for ADLs. Compared to RGB videos, depth video-based activity recognition is less intrusive and eliminates many privacy concerns, which are crucial for applications such as life-logging and ambient assisted living systems. Existing methods rely on handcrafted features for depth video classification and ignore the importance of audio stream. In this paper, we propose an ADL recognition system that relies on both audio and depth modalities. We propose to adopt popular convolutional neural network (CNN) architectures used for RGB video analysis to classify depth videos. The adaption poses two challenges: (1) depth data are much nosier and (2) our depth dataset is much smaller compared RGB video datasets. To tackle those challenges, we extract silhouettes from depth data prior to model training and alter deep networks to be shallower. As per our knowledge, we used CNN to segment silhouettes from depth images and fused depth data with audio data to recognize ADLs for the first time. We further extended the proposed techniques to build a real-time ADL recognition system.