FMCW Radar Fall Detection Based on Parallel Multi-scale Feature Extraction

Existing fall detection systems are prone to insufficiency of feature extraction, over-fitting and error propagation caused by small datasets, as well as the low recognition rate, this paper proposes a RNCK-SENet model for fall detection. Firstly, an imbalanced human motion echo signal is obtained using FMCW radar, which is then utilized to construct a set of micro-Doppler feature maps. Next, preliminary feature extraction is performed using a transfer learning-based ResNet18 model, In addition, the features are filtered and optimized through the SE module to fully extract comprehensive behavioral features. To capture information from multiple scales and enhance feature extraction, a parallel multi-scale feature extraction module is embedded in the network. Finally, the fused features are fed into a fully connected layer for classification. Comparative results with classical networks demonstrate that our fall detection algorithm achieves higher average accuracy, recall, and F1 score, reaches 98.43%, 99.33%, and 99.12% respectively. The F1 score is improved by 2% to 10%.