DeepConvLSTM on single accelerometer locomotion recognition

Abstract: This project aims to evaluate the deep neural network architecture Deep-ConvLSTM to classify locomotive human activities using data from a single accelerometer. The evaluation involves comparisons to a simpler convolutional neural network and a hyperparameter evaluation in regards to the networks number of convolutional layers. The benchmark OPPORTUNITY dataset is used for training and evaluation from which triaxial accelerometer data from hips and legs are extracted. The results of the evaluation suggests that DeepConvLSTM outperforms simpler models on most locomotive activity recognition, especially at filtering out unclassified data. Further the results show that DeepConvLSTMs performance improves with a higher number of convolutional layer, but the number of limited by the architectures lack of padding and is compensated by longer training times.