Spak-Net: A Lightweight Convolutional Neural Network for Activity Recognition with Wearable Inertial Sensors

The prevailing Internet of Health and Medical Things (IoHMT) strategy emphasizes preventing disease onset by continuously monitoring individuals’ physical activities. This approach makes Human Activity Recognition (HAR) and Behavior Analysis crucial areas of research in the field of IoHMT. A specific subset of HAR, known as Sports Activity Recognition (SAcR), aims to detect and classify sports movements. Three primary methodologies are used to monitor such activities: computer vision, environmental sensors, and wearable sensors. Among these, wearable sensors emerge as the most viable solution, based on an assessment of the advantages and limitations of each technology. This paper introduces Spak-Net, a model characterized by its efficient u se o f a ddition l ayers a nd c ontaining o nly 398 trainable parameters. Leveraging inertial sensor data from wearable devices, the model demonstrated exceptional performance, achieving 98.93% ± 0.9% (95% CI) accuracy on the sports activity recognition IM-Sporting Behavior dataset and 96.25% ± 1.2% (95% CI) accuracy on the human activity recognition WISDM-11 dataset making it the most efficient n eural network for achieving this much performance on WISDM11 dataset to the best of our knowledge.