Deep learning fusion conceptual frameworks for complex human activity recognition using mobile and wearable sensors

Abstract

Human Activity recognition (HAR) is an important area of research in ubiquitous computing and Human Computer Interaction. To recognize activities using mobile or wearable sensor, data are collected using appropriate sensors, segmented, needed features extracted and activities categories using discriminative models (SVM, HMM, MLP etc.). Feature extraction is an important stage as it helps to reduce computation time and ensure enhanced recognition accuracy. Earlier researches have used statistical features which require domain expert and handcrafted features. However, the advent of deep learning that extracts salient features from raw sensor data and has provided high performance in computer vision, speech and image recognition. Based on the recent advances recorded in deep learning for human activity recognition, we briefly reviewed the different deep learning methods for human activities implemented recently and then propose a conceptual deep learning frameworks that can be used to extract global features that model the temporal dependencies using Gated Recurrent Units. The framework when implemented would comprise of seven convolutional layer, two Gated recurrent unit and Support Vector Machine (SVM) layer for classification of activity details. The proposed technique is still under development and will be evaluated with benchmarked datasets and compared with other baseline deep learning algorithms.