Falls in wet bathroom environments are a major safety risk for seniors living alone. Recent work has shown that mmWave-only, vibration-only, and existing multimodal schemes, such as vibration-triggered radar activation, early feature concatenation, and decision-level score fusion, can support privacy-preserving, non-intrusive fall detection. However, these designs still treat motion and impact as loosely coupled streams, depending on coarse temporal alignment and amplitude thresholds, and do not explicitly encode the causal link between radar-observed collapse and floor impact or address timing drift, object drop confounders, and latency and energy constraints on low-power edge devices. To this end, we propose a two-stream architecture that encodes radar signals with a Motion--Mamba branch for long-range motion patterns and processes floor vibration with an Impact--Griffin branch that emphasizes impact transients and cross-axis coupling. Cross-conditioned fusion uses low-rank bilinear interaction and a Switch--MoE head to align motion and impact tokens and suppress object-drop confounders. The model keeps inference cost suitable for real-time execution on a Raspberry Pi 4B gateway. We construct a bathroom fall detection benchmark dataset with frame-level annotations, comprising more than 3~h of synchronized mmWave radar and triaxial vibration recordings across eight scenarios under running water, together with subject-independent training, validation, and test splits. On the test split, our model attains 96.1% accuracy, 94.8% precision, 88.0% recall, a 91.1% macro F1 score, and an AUC of 0.968. Compared with the strongest baseline, it improves accuracy by 2.0 percentage points and fall recall by 1.3 percentage points, while reducing latency from 35.9 ms to 15.8 ms and lowering energy per 2.56 s window from 14200 mJ to 10750 mJ on the Raspberry Pi 4B gateway.
Figures & tables
Fig. 1: Overall architecture of the proposed fall detection framework. Radar signals are processed by the Motion–Mamba stream and vibration signals by the Impact–Griffin stream. The two branches are fused through cross–conditioned interaction, low–rank bilinear coupling, and a fusion Switch–MoE.
Fig. 2: Experimental bathroom mock-up ( 3.70×2.50m ). Left: top-down floor plan showing the shower bay, fixtures, and reference dimensions. Right: front view illustrating the sink, mirror, partition, and furniture placement.
Fig. 3: Site photograph and hardware. Left: C4001 radar and ADXL345 node electronics (ESP32-C3 MCU). Right: installed locations—radar on the upper wall ( 1.87m height), vibration node on the shower platform ( 0.18m above floor); annotated room dimensions.
Scenario
Eval. Windows (N)
Non–Fall Metrics (%)
Fall Metrics (%)
Window-level (%)
Precision
Recall
Precision
Recall
Acc.
Bal. Acc.
Macro F1
Per-scenario performance
Empty Bathroom
218
97.00
92.50
96.20
88.10
97.10
90.30
93.33
Light Object Drop
498
94.80
93.00
92.00
87.60
96.38
90.30
91.80
Heavy Object Drop
380
90.10
77.50
85.00
64.40
91.30
70.95
78.30
Normal Walking
1188
97.20
95.60
96.10
91.50
97.83
93.55
95.08
TABLE I: Scenario-wise evaluation in the wet bathroom environment. For each scenario we report per-class precision and recall for Non–Fall and Fall windows and window-level Accuracy, Balanced Accuracy and Macro F1. The bottom block summarizes robustness tests of the full model under synthetic time misalignment and modality dropouts across all scenarios.
Fig. 4: Eight background scenarios and a representative intentional fall in the wet bathroom environment: (a) empty scene; (b) soap drop; (c) mop drop; (d) upright walking; (e) flexed‐torso walking; (f) wall‐assisted walking; (g) quiet standing; (h) squatting; (i) example of an intentional fall. All human activities were performed with the shower running.
Fig. 5: Scenario-wise precision–recall curves (left) and overall ROC curve (right). The micro-averaged ROC yields AUC=0.968 which indicates high true-positive rate at low false-positive rate.
Subsystem
Hardware
Configuration
Params [M]
MACs [M]
Mod. size [MB]
Latency [ms]
Pavg [W]
Ewin [mJ]
Cap. [win/s]
Sensing node
ESP32-C3 + ADXL345
Vibration preprocessing
–
–
–
pre inf end 0.9 – 0.9
0.32
820
–
ESP32-C3 + C4001
Radar front end
–
–
–
pre inf end 2.1 – 2.1
1.85
4740
–
Edge gateway
Raspberry Pi 4B
radar and vibration fusion
2.1
310
8.4
pre inf end 8.9 6.7 15.8
4.20
10750
63.3
Raspberry Pi 4B
Motion–Mamba radar only
1.3
210
5.2
pre inf end 7.5 19.8 28.4
3.80
9730
35.2
Raspberry Pi 4B
Impact–Griffin vibration only
0.9
140
3.6
pre inf end 4.1 13.2 19.1
3.50
8960
52.4
TABLE II: Latency and energy per 2.56 s window for sensing nodes and inference back ends. The latency column lists preprocessing, model inference, and end-to-end detection delay in milliseconds. Power and energy are averaged over 200 windows in the wet bathroom environment.
Variant
Architectural components
Modalities
Metrics on test set (%)
LSK1D
LR-Temp.
ICA
Stream-MoE
X-Cond.
MLB
Fusion-MoE
Vib.
Radar
Acc.
Prec.
Rec.
Macro F1
AUC
Single modality baselines
Vibration only
✓
✓
✓
✓
✓
92.1
88.7
81.4
84.9
94.5
Radar only
✓
✓
✓
✓
86.7
74.3
89.5
81.2
93.4
Fusion baselines
Early concatenation with CNN head
✓
✓
✓
92.0
91.0
82.5
86.5
95.2
TABLE III: Ablation of architectural components and fusion strategies on the test split. ✓ indicates that the component or modality is enabled. All models are trained with the same protocol on the same train and validation partitions. Metrics are reported as percentages and averaged over three random seeds.
Method
Modalities
Acc. [%]
Fall Rec. [%]
Lat. [ms]
Ewin [mJ]
Thresh. accel [ 27 ]
Vibration only
88.3
79.5
22.7
14931
ML wearable [ 28 ]
Vibration only
90.6
83.1
34.8
19210
vibration rule [ 19 ]
Vibration only
89.1
80.2
25.1
12131
vibration ML [ 5 ]
Vibration + mmWave
91.4
84.0
17.6
11153
mmWave 1D CNN [ 6 ]
Radar only
92.0
86.1
17.9
9800
FMCW LSTM Transf. [ 7 ]
Radar only
93.2
87.3
28.7
12100
TABLE IV: Comparison with representative fall detection methods reimplemented on our dataset. Window metrics and energy are averaged on the Raspberry Pi 4B gateway.