Optimizing Memory Footprint for Radar-Based Human Activity Recognition on Resource-Constrained Devices

Abstract

Human activity recognition is useful for tasks like patient fall detection. While radar offers privacy-preserving sensing, deploying predictive networks on edge devices faces challenges due to their memory requirements. Previous attempts to address this through hyperparameter tuning often lead to unwanted accuracy decline. On the other hand, pruning is used to reduce memory requirements of deep neural networks, but its effectiveness is limited for compact neural networks such as the Split BiRNN. The Split BiRNN architecture segregates the forward and backward computations across different components enabling quick initial predictions on the edge device and a refined prediction executed in the cloud. Our study enhances a prominent pruning algorithm to reduce memory requirements, making both models more compact. This not only reduces hardware costs but also enhances evaluation speed. Our findings suggest that pruning can sufficiently compact the network while maintaining its accuracy.