Radar-Based Human Activity Recognition Using Optimized CNN on Edge Devices

Human Activity Recognition (HAR) has a significant impact on ensuring safety surveillance and health care. The Convolutional Neural Networks (CNNs) have gained popularity in classifying human activities based on micro-Doppler signatures. However, the large amount of parameters within the CNN model results in elevated computational expenses and a larger model size. Resource limitations in edge devices often constrain the deployment of radar-based HAR with deep learning models. This study proposes an optimized CNN-based deep learning model by utilizing transfer learning, fine-tuning, and quantization techniques. The proposed model is evaluated on a public HAR dataset and compared with other models. Experimental results show that the proposed model yields a substantial reduction in model size (~4 times smaller) while preserving high performance (accuracy of 97.71 % and F1 score of 97.85%). The average inference time on Raspberry Pi 4 + Google Coral USB accelerator edge device is 0.07 s. The overall results suggest that the proposed model is suitable for radar-based HAR on resource-constrained edge devices.