Research map / Detection and segmentation / CNN features

WiFi CSI Based Temporal Activity Detection via Dual Pyramid Network

AAAI 2025 · Detection and segmentation › CNN features · approach: Context for hard detection conditions · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DPWiT, a dual pyramid network with frequency-split temporal encoders and a learning-free fluctuation encoder for WiFi CSI temporal activity detection, plus a new untrimmed CSI dataset.

Problem

Temporal activity detection on long, untrimmed WiFi CSI signals, where activity boundaries are unknown and vision TAD methods transfer poorly to noisy wireless data.

Approach

Split feature learning into a low-frequency transformer branch with Signed Mask-Attention and a high-frequency conv-pool branch fused by ContraNorm, add a learning-free max-min fluctuation encoder, and fuse the two feature pyramids with cross-attention.

Key results

• DPWiT reaches 74.5 average mAP (85.5 at tIoU 0.3) on the new CSI dataset, versus 58.0 for the best baseline DyFADet.
• Compared with DyFADet, DPWiT uses 44.1 GFLOPs versus 304.0 GFLOPs, though its inference time is 61.8 ms versus 106.0 ms.
• Ablation: replacing Signed Mask-Attention with standard self-attention drops average mAP from 74.5 to 59.4.

Measured effects reported in the paper

Surprising findings

• In preliminary tests, the convolutional network beat the transformer on localization mIoU (49.4 vs 47.1), while the transformer beat it on classification precision (21.9 vs 18.3).

Limitations

• Evaluation is on a single self-collected dataset with three volunteers in one office room.
• Only the dataset split 7:3 is reported; cross-person evaluation is deferred to supplementary material.

Experimental setup

Datasets: WiFi CSI untrimmed activity dataset (self-collected)

Benchmarks: self-collected WiFi CSI untrimmed activity dataset

Metrics: mAP at tIoU 0.3-0.7 mAP avg GFLOPs inference time false positive error breakdown

Baselines: ResNet1d THAT AFSD BREM ActionFormer TADTR Tridet TemporalMaxer DyFADet

Builds on: ActionFormer ContraNorm DyFADet Soft-NMS

Closest papers by meaning

• CSI-Bench: A Large-Scale In-the-Wild Dataset for Multi-task WiFi Sensing (NeurIPS 2025 · similarity 0.88)
• Two-Stream Convolution Augmented Transformer for Human Activity Recognition (AAAI 2021 · similarity 0.88)
• ConSense: Continually Sensing Human Activity with WiFi via Growing and Picking (AAAI 2025 · similarity 0.87)
• Decomposing and Fusing Intra- and Inter-Sensor Spatio-Temporal Signal for Multi-Sensor Wearable Human Activity Recognition (AAAI 2025 · similarity 0.80)
• HARDVS: Revisiting Human Activity Recognition with Dynamic Vision Sensors (AAAI 2024 · similarity 0.79)
• DiGIT: Multi-Dilated Gated Encoder and Central-Adjacent Region Integrated Decoder for Temporal Action Detection Transformer (CVPR 2025 · similarity 0.79)