A Real-Time Deployable Attention-Driven CNN–LSTM Framework for Human Activity Recognition Using Wearable Sensor

Human activity recognition (HAR) using wearable inertial sensors is increasingly vital for healthcare, rehabilitation, and pervasive computing. Yet, many state-of-the-art (SoTA) models rely on handcrafted feature engineering, heavy time–frequency transformations, or large neural networks, limiting real-time deployment on resource-constrained smartphones and wearables. We propose an end-to-end, lightweight attention-guided CNN–CBAM–LSTM architecture that directly consumes raw IMU signals standardized by StandardScaler, eliminating costly preprocessing. Our design uniquely combines: 1) multiresolution convolutional kernels to capture diverse motion dynamics; 2) convolutional block attention modules (CBAMs) to adaptively highlight salient spatial features and suppress sensor noise; and 3) temporal self-attention atop stacked long short-term memory (LSTM) layers to model long-range temporal dependencies. Extensive experiments on three public HAR benchmarks—mHealth (controlled), MotionSense (uncontrolled), and UCI HAR (semi-controlled)—show that our model achieves accuracies of 99.7%, 99.5%, and 99.4%, respectively, consistently surpassing recent baselines. Moreover, structured pruning and quantization compress the model to just 130.8 KB, while real-world tests on a Snapdragon 845-class Android device confirm ultralow inference latency of 13.5 ms per segment under live streaming. This fusion of high accuracy, low latency, and minimal memory footprint demonstrates the potential of attention-driven, multiresolution deep learning for practical, real-time edge AI in HAR applications.