An Improved Deep Convolutional LSTM for Human Activity Recognition Using Wearable Sensors

Behavior recognition technologies based on sensor devices show strong development potential as sensor devices are increasingly applied in people’s lives. However, some challenges in sensor-based behavior recognition, such as incomplete feature extraction, complex behavior recognition algorithms, and the recognition accuracy, remain open issues. To address these issues, this article proposes an improved deep convolutional long short-term memory (DeepConvLSTM) model that uses the squeeze and excitation (SE) module and group convolutions to perform human activity recognition (HAR). The proposed scheme uses convolutional neural network (CNN) and LSTM to simultaneously extract temporal and spatial features. The SE module is an attention mechanism that allows the network to perform feature recalibration, through which the network learns to use global information to selectively emphasize informative features while suppressing less useful features. Moreover, the model uses group convolutions to optimize the number of parameters and reduce the computational complexity. To show the benefit of the proposed model, the performance and efficiency of the improved DeepConvLSTM model and other traditional models are compared on the OPPORTUNITY dataset. In addition, this article evaluates the impact of hyperparameters on the proposed model and designs an ablation experiment to verify the improvement of each modification. The evaluation results show that the improved DeepConvLSTM model increases the HAR accuracy up to 90.80% given that most computational resources are assigned to more informative features. Moreover, floating-point operations per second (FLOPs) are reduced by 88.93% when the number of group is set to 16 because the group convolution reduces the training parameters greatly.