Hybrid classification framework for human activity recognition using madgwick sensor fusion and lightweight deep learning
摘要
Human Activity Recognition (HAR) is a key field at the intersection of the Internet of Things and Machine Learning (ML). It involves identifying usual activities such as walking or standing, and has critical applications in domains like healthcare, monitoring, and robotics. Achieving high performance in HAR requires integration of data from multiple sensors, including the accelerometer, magnetometer, and gyroscope. However, most existing works rely on naive sensor fusion via simple concatenation that often overlooks the benefits of advanced fusion techniques necessary to enrich contextual information for HAR models. To address this limitation, we propose a novel approach using the Madgwick algorithm to fuse data from wearable sensors for HAR. Our approach first estimates orientation using the Madgwick algorithm, which employs a quaternion representation, and then updates this estimate based on data from the accelerometer and magnetometer. Second, it integrates Euler angles for better contextualization and finally projects the accelerometer data into a global reference frame. Concurrently, we introduce a lightweight deep learning (DL) model for automatic feature extraction from the fused data by combining depthwise convolutions and an attention mechanism. For classification, we explore two strategies: one based on a Fully Connected Network (FCN) and the other using the classical ML model Extremely Randomized Trees (ETR). To validate the proposed approach, we conducted experiments on two public HAR datasets, PAMAP2 and DOMINO, and reported the results using several metrics. The results demonstrate the effectiveness of our approach, showing an improvement in the