Toward UAV-assisted Wi-Fi sensing via deep neural networks: a benchmark study on activity and gesture recognition

Purpose Wi-Fi channel state information (CSI) sensing is a candidate technology for future Unmanned Aerial Vehicle (UAV)-assisted, device-free human sensing. This paper benchmarks convolutional neural network (CNN) architectures for CSI-based activity and gesture recognition under a controlled, repeated protocol, as a first step towards selecting models for resource-constrained platforms. Design/methodology/approach Two public CSI datasets collected with fixed indoor Wi-Fi infrastructure were used: UT-HAR (activity recognition, 7 classes) and Widar (gesture recognition, 22 classes, body-coordinate velocity profile). Two VGG-inspired CSI-adapted networks (CSI-VGGNet-S and CSI-VGGNet-D), GoogLeNet and DenseNet were compared with four retrained SenseFi baselines (MLP, ResNet-18, LSTM and CNN-GRU). All models were trained with three random seeds under one common budget (at most 20 epochs, early stopping on a held-out validation partition), and the test partitions were used once per run. Sliding-window overlap in UT-HAR was quantified, and recording-level (UT-HAR) and cross-user (Widar) splits were used to probe generalization. No UAV hardware or UAV-collected CSI was used. Findings CSI-VGGNet-D achieved the highest mean test accuracy on UT-HAR (96.80 ± 0.87%) and on Widar (74.43 ± 1.31%). Depth alone did not predict accuracy: GoogLeNet and DenseNet varied strongly across seeds on UT-HAR (SD 16.1 and 12.8 percentage points) despite high training accuracy, and model rankings differed between the two tasks. In the released UT-HAR split, 92% of test windows share at least half of their raw frames with a training window; on a recording-level split without shared frames, the accuracy of CSI-VGGNet-D fell by 6.6 percentage points. Released-split accuracies therefore overstate generalization to unseen recordings, and accuracy, model size and latency must be weighed jointly. Research limitations/implications Results come from fixed indoor infrastructure and a desktop CPU; single-thread latency is only a proxy for embedded hardware, and three seeds give limited statistical power. Validation with UAV-collected CSI and onboard hardware is required before any claim about UAV deployment can be made. Originality/value The paper does not propose a new deep-learning architecture. Its contribution is empirical: a multi-seed, validation-controlled, matched-protocol comparison that re-evaluates SenseFi baselines alongside VGG-style, inception and DenseNet CNNs, quantifies train-test window overlap in the widely used UT-HAR split and its effect on reported accuracy and reports accuracy together with parameters, FLOPs and latency as input to future UAV-oriented model selection.