A Swarm Intelligence Assisted IoT-Based Activity Recognition System for Basketball Rookies

Recent years have witnessed many applications of wearable sensor technology and machine learning on smart sports training, e.g., basketball. For rookie players, the fundamental skills are very important and can be improved using a scientific auxiliary training system, the core of which is human activity recognition (HAR) technique. For basketball players, one complete action is usually performed in a moment with significant dynamics and high-frequency spikes or noise. So, extracting meaningful features from intricate sensor signals is one of the important prerequisites. Besides, to improve the recognition accuracy, it is challenging to empirically determine an ideal feature subset out of the extracted high-dimensional features for particular types of basketball activities, which is essentially an NP-hard optimization problem. To address the above issues, we propose a smart activity recognition system hybridizing an IoT-based edge-cloud system and a swarm intelligence-based feature selection algorithm, named Adaptive Binary Particle Swarm Optimization (ABPSO). The devices can automatically collect and process basketball players' action signals. A comprehensive feature set with 300 features that can cover and capture the characteristics of different activities is established. Then, ABPSO with global search ability is able to identify the optimal feature subset combined with classifiers for accurate activity recognition. Thanks to ABPSO, the size of machine learning model and the amount of transmitted data is largely reduced, which makes it well compatible for IoT-based applications. Experiments performed on nine human subjects demonstrate that ABPSO can achieve accuracy of 97.26% for five fundamental activities on average, which outperforms five state-of-the-art FS algorithms, two classic filter-based FS methods and one deep learning based method.