Accurate Fitness Activity Recognition Based on IMU Sensors Using Deep Residual Bidirectional Gated Recurrent Unit Network

This study presents an accurate and efficient approach to fitness activity recognition by combining inertial measurement unit (IMU) sensor data with deep learning techniques. We propose a novel architecture, 1D-CNN-ResBiGRU, which integrates one-dimensional convolutional neural networks (1D-CNNs) for spatial feature extraction with residual bidirectional gated recurrent units (ResBiGRUs) for temporal sequence modeling, enabling precise classification of diverse fitness activities. The ResBiGRU introduces residual connections within the bidirectional gated recurrent units, not just between layers. The effectiveness of the proposed framework was validated using the publicly available Running Exercise dataset, which comprises IMU recordings from 20 participants performing seven distinct running exercises. Experimental results demonstrate that the model achieves a classification accuracy of 99.92% with a one-second window, substantially outperforming traditional machine learning methods and existing deep learning architectures. The inclusion of residual connections within the BiGRU network significantly enhances performance by improving gradient flow and capturing complex temporal dependencies. Further evaluations under challenging conditions—such as reduced sensor configurations and shortened time windows—confirm the model’s robustness. Notably, the model maintains a high accuracy of 99.80% even with a minimal 0.25-second window, underscoring its suitability for real-time applications with low latency requirements. This work advances IMU-based fitness activity recognition by demonstrating that integrating residual structures with bidirectional recurrent models can significantly improve recognition performance. The findings contribute to the development of intelligent mobile fitness applications that provide personalized, responsive exercise feedback and performance monitoring.