Random Search One Dimensional CNN for Human Activity Recognition

Due to its wide application, human activity recognition (HAR) has become a common subject for research specially with the development of deep learning. Many researchers believe that deep convolutional neural networks (DCNN) are ideal for feature extraction from signal inputs. This has gained widespread interest in using these methods to identify human actions on the mobile phone in real time. A deep network architecture using random search one dimensional convolutional neural network (RS-1D-CNN) is proposed to find best networks connections and hyper-parameters to enhance model performance. Batch normalization (BN) layer was added to speed up the convergence. Moreover, we have applied a global average pooling (GAP) for dimensionality reduction and to reduce model hyper-parameters, followed two dense connected layers. The final dense layer has a softmax activation function and a node for each potential object category. Public UCI-HAR dataset was used to evaluate model performance. Random search has been utilized to perform hyper parameter tuning to determine the optimal model parameters. Proposed model will automatically extract and classify human behaviours. Daily human activities that provided by UCI-HAR include (walking, jogging, sitting, standing, upstairs and downstairs). Results has shown that our approach outperforms both CNN, LSTM method and other state-of-the-art approaches.