IoT based human activity recognition on drifted data stream using arbitrary width convolution neural network

A common research focus in deep learning is human activity recognition (HAR), which involves detecting human activities using sensor data from magnetometers, accelerometers, and gyroscopes. For real-time HAR applications, it's crucial to develop a model that is both cost-effective and efficient in terms of resource and processing power usage. To achieve this, our approach trains the deep learning model on channels of variable width instead of adjusting the number of neurons or layers. To reduce computational overhead, random sampling is applied to the lower triangular convolution layer. The model detects human activity from streaming sensor data using adaptive window sizes, which are designed to address sudden changes in activity, known as drifts, such as falls or collapses. The adaptive window strategy is a key to manage dynamic window sizes and handle drifts effectively. The model's usability and practicality are evaluated on a range of IoT devices and tested on five real-world datasets, as well as one synthetic dataset generated in real-time using a Raspberry Pi 3B and a NodeMCU. Experimental results show that our model achieves a higher accuracy of 97.84% on the WISDM dataset with a width of 0.85, outperforming other state-of-the-art methods.