Optimized Data Transmission for Radar-Based Edge-Cloud Human Activity Recognition via Quantization

Abstract

Radar systems offer a privacy-preserving approach to human activity recognition, leveraging Deep Neural Networks (DNNs) for effective data processing. However, the scalability of large DNNs becomes challenging when handling multiple radar streams simultaneously, such as monitoring several rooms in a hospital. A viable solution, the Split Bidirectional Recurrent Neural Network (Split BiRNN) for Human Activity Recognition, addresses this issue by dividing data processing between on-premise (low-power, low-cost devices) and off-premise (higher performance devices) components. This paper delves into the optimization of data transmission between these two components, considering the significant battery consumption of edge devices due to data transmission. Through quantization techniques, the paper explores the reduction of data size while maintaining accuracy levels. Results indicate that quantization to 2 or 1 bits is feasible, with minimal accuracy loss of 0.21% and 1.12%, respectively.