Convolutional Neural Network with CBAM Module for Fitness Activity Recognition Using Wearable IMU Sensors
Sakorn Mekruksavanich, Ponnipa Jantawong, Wikanda Phaphan, Anuchit Jitpattanakul · 2024
Wearable motion sensors are gaining traction for fitness activity tracking due to their utility in assisted living, wellness tracking, and fitness training applications. Such sensors typically incorporate accelerometers and gyroscopes to monitor and log physical activities. This work proposes a deep-learning approach for identifying fitness activities using wearable inertial measurement units (IMUs) signals. Specifically, a convolutional neural network (CNN) integrated with a convolutional block attention module (CBAM) provides the core framework. The CBAM module applies channel and spatial attention mechanisms to enhance CNN’s capacity to discriminate salient features. Multichannel IMU time-series data, comprising accelerometer and gyroscope readings as subjects perform various exercises, supply input to the model. By learning via an attention-enhanced architecture, the model can extract robust and informative representations to categorize activities accurately. Experiments on a public IMU dataset with five exercise types across over twenty participants evaluated the approach. Results showed over 99.58% subject-independent activity recognition accuracy when augmenting the CNN with CBAM, constituting significant performance gains. The research findings indicate that an efficient deep-learning solution with attention modeling can recognize fitness activities reliably from on-body sensors. The proposed methodology holds substantial promise for applications in sports analytics and personalized health tracking.