GKR-Spike: A Cross-Attention Enhanced Hybrid SNN via Gaussian Kernel Density Estimation and Key-Residual Frames for AIoT

Abstract

Device-free fall detection using Wi-Fi Channel State Information (CSI) is promising for pervasive health monitoring but is impeded by environmental dependency and a high computational cost. While deep learning (DL) methods can automatically extract features, computationally expensive 3D Convolutional Neural Networks (CNNs) restrain their deployment on resource-constrained edge IoT devices. To strike a balance between system robustness and computational efficiency, we propose GKR-Spike, an end-to-end cross-environment fall detection system. GKR-Spike utilizes Gaussian Kernel Density Estimation (GKDE) to transform a one-dimensional CSI amplitude time series into statistical probability-density heatmaps. Then, GKDE-encoded distribution inputs are decomposed into spatial key frames and temporal residual frames to separate segment-level spatial context from motion-related dynamics. Based on these complementary inputs, GKR-Spike adopts a low-power hybrid SNN-ANN architecture built around two Spiking ResNet-18 feature streams. Sparse spike-driven synaptic computation can replace many dense multiply–accumulate (MAC) operations with conditional accumulations, offering the potential to reduce operation energy. Furthermore, we introduce a multi-scale cross-attention mechanism for effective feature fusion between the dual spiking streams. Experimental evaluations demonstrate that, without target-domain model retraining, GKR-Spike achieves 92.52% Accuracy, 92.35% Balanced Accuracy, and 92.14% Recall in Lab under transductive calibration using unlabeled target scores. In the compact, furniture-filled Room, it achieves 86.01% Accuracy, 79.96% Balanced Accuracy, 72.41% Recall, PR-AUC 0.4958, and a 12.50% false positive rate. The estimated operation energy is 16.49–16.72 mJ per sample; measured batch-one inference on an RTX 4090 D GPU has a mean latency of 45.36 ms and a total peak allocation of 160.28 MiB.