Search papers, labs, and topics across Lattice.
This paper introduces a quantized deep neural network (QDNN) framework for real-time human activity recognition (HAR) on edge devices, addressing resource constraints through weight pruning, fixed-point quantization, and layer compression. The framework incorporates an adaptive quantization-aware training strategy to maintain accuracy during quantization and integrates optical sensors for enhanced data acquisition. Experimental results on UCI HAR and WISDM datasets using Raspberry Pi and NVIDIA Jetson Nano show significant improvements in memory consumption, inference latency, and computational complexity while preserving high classification accuracy.
Achieve >98% accuracy in human activity recognition on edge devices while slashing memory consumption by 45% and inference latency by 38% using quantized deep neural networks.
Quantized deep neural networks (QDNNs) provide an efficient solution for real-time human activity recognition (HAR) on resource-constrained edge devices by addressing computational and memory limitations while maintaining high classification accuracy. This study proposes a QDNN-based framework utilizing advanced quantization techniques, including weight pruning, fixed-point quantization, and layer compression, to optimize model size, inference speed, and resource efficiency. Additionally, the integration of optical sensors for high-resolution data acquisition further enhances the robustness and sensitivity of activity recognition systems. An adaptive quantization-aware training strategy is integrated into the framework to mitigate accuracy loss caused by quantization and ensure robust performance in dynamic environments. The framework was evaluated on benchmark datasets, UCI HAR and WISDM, using Raspberry Pi and NVIDIA Jetson Nano as deployment platforms. Experimental results demonstrate a 45% reduction in memory consumption, a 38% decrease in inference latency, and up to a 42% reduction in computational complexity compared to baseline DNNs, while maintaining over 98% classification accuracy across various quantization levels. These results highlight the practicality of QDNNs for enabling real-time HAR in wearable health monitoring, smart homes, and other edge-based applications, offering a scalable and efficient approach to deploying machine learning models in resource-limited settings.
