Published Paper Details:

This paper presents the design of a multiscale convolutional neural network (MCNN) for the task of human behavior recognition, with the aim of increasing accuracy and reducing complexity of the model. Previous models such as CNN2D and LSTM time series modeling relied on a mean global average but neglected other spatial and depth features resulting in inaccuracies. The MCNN model is based on the concept of space-time interaction and depth-separable convolution modules inserted into a CNN3D model in which both the spatial and temporal information are enhanced. The system was trained and evaluated on the UCI HAR dataset which has six activity labels that were recorded using smartphone's sensors. The evaluation showed that the new model offered a 94% accuracy rate while improving learning complexity. An extended hybrid model comprising of a combination of convolutional neural networks, Gated recurrent units, and bidirectional algorithms produced an accuracy of 96% while using less parameters thus showing tremendous effectiveness and capability for real world tasks.