COSMOS: Networked ISAC Enabled Target Recognition Towards Low-Altitude Economy

Hongliang Luo¹ Chuanbin Zhao¹ Boxuan Sun¹ Zhonghua Chu¹ Shengjie Quan¹

Guangyi Liu² Feifei Gao¹

¹Tsinghua University ²China Mobile Research Institute

ABSTRACT

In this paper, we propose a low-altitude target (LAT) recognition scheme based on multi-base station (BS) collaboration and multi-scale feature fusion for integrated sensing and communications (ISAC) network. We formulate the motion equations, echo channels, and echo signals for unmanned aerial vehicle (UAV), bird, vehicle, and pedestrian under multi-BS collaborative monitoring scenario. Then we extract the velocity-resolution-preferred time-frequency spectrum, time-resolution-preferred time-frequency spectrum, and velocity-transfer time-frequency spectrum observed by each BS from echo signals. We collectively refer to these three types of time-frequency spectrum as the multi-scale feature of the LAT. Next, we design a multi-BS and multi-scale feature fusion enabled LAT recognition network with Swin Transformer, which employs the visualized images of multi-scale feature to jointly recognize the target through deep feature extraction, intra-BS feature interaction, inter-BS feature interaction, and target recognition output. We generate a massive echo signal dataset comprising 1,440,000 samples for LAT recognition within ISAC network. This dataset can serve as a public benchmark to evaluate our proposed scheme and facilitate future research. Simulation results demonstrate that the proposed scheme realizes high recognition accuracy and robust unseen-subtype generalization, confirming the effectiveness of multi-scale feature fusion and the additional gains brought by multi-BS collaboration.

Multi-BS scenario

Networked ISAC sensing with three cooperative base stations.

Multi-scale features

VRP-TF, TRP-TF, and VT-TF capture complementary motion cues.

Hierarchical fusion

Shared Swin-B with intra-BS and inter-BS feature interaction.

Large-scale dataset

1,440,000 samples across 40 target subtypes and six SNR settings.

System Model

The recognition task is built on a networked ISAC sensing cell, where three BSs observe low-altitude targets from complementary geometric views.

LAT recognition scenario in an ISAC network with three base stations and four target categories — **Fig. 1.** LAT recognition scenario in ISAC network.

Aerial view of multi-BS cooperative sensing scenario with BS-A, BS-B, BS-C and target coordinates — **Fig. 2.** Aerial view of multi-BS cooperative sensing scenario.

Multi-Scale Features

VRP-TF spectrum

Velocity-Resolution-Preferred Time-Frequency Spectrum

VRP-TF is obtained by using a long STFT window, which provides high velocity resolution and better observation of LAT velocity details. The resulting spectra show class-dependent horizontal stripe patterns.

TRP-TF spectrum

Time-Resolution-Preferred Time-Frequency Spectrum

TRP-TF is obtained by using a short STFT window, which preserves high time resolution. It displays periodic envelopes caused by the periodic micro-motions of different LAT categories.

VT-TF spectrum

Velocity-Transfer Time-Frequency Spectrum

VT-TF replaces the velocity-FFT calculation with matched filtering and is composed of velocity component transition amounts. It reflects micro-motion periodicity while retaining more velocity distribution details.

Low-Altitude Target Recognition Network

We design a multi-BS and multi-scale feature fusion enabled LAT recognition network with Swin Transformer, which employs the visualized images of multi-scale feature to jointly recognize the target through deep feature extraction, intra-BS feature interaction, inter-BS feature interaction, and target recognition output.

Swin-B based intra-BS and inter-BS fusion network

Large-Scale Dataset

4 target categories

40 target subtypes

240,000 motion models

6 SNR settings

1,440,000 echo signal samples

Experimental Results

Experimental Settings

Seen-subtype setting

Training and testing samples share the same target subtypes, evaluating recognition accuracy under matched subtype distributions.

Unseen-subtype setting

Testing samples contain target subtypes excluded from training, evaluating generalization to new low-altitude target subtypes.

Performance Summary

99.88% average accuracy on seen subtypes

97.82% average accuracy on unseen subtypes

96.08% accuracy at 3 dB on unseen subtypes

Seen-subtype recognition accuracy (%)

Scheme	3 dB	8 dB	13 dB	18 dB	23 dB	No noise	Avg.
BS-A: VRP	94.70	97.69	98.79	99.00	99.14	99.11	98.07
BS-A: TRP	95.51	98.50	99.20	99.43	99.39	99.44	98.58
BS-A: VT	93.01	97.63	98.88	99.16	99.30	99.32	97.88
BS-A: VRP+TRP+VT	96.52	98.88	99.42	99.57	99.63	99.65	98.94
BS-B: VRP+TRP+VT	97.00	98.84	99.33	99.52	99.50	99.53	98.95
BS-C: VRP+TRP+VT	96.75	98.81	99.38	99.52	99.53	99.55	98.92
BS-A/B/C: VRP+TRP+VT	99.52	99.91	99.95	99.97	99.97	99.97	99.88

Unseen-subtype recognition accuracy (%)

Scheme	3 dB	8 dB	13 dB	18 dB	23 dB	No noise	Avg.
BS-A: VRP	87.64	90.03	91.06	91.53	91.74	92.14	90.69
BS-A: TRP	90.05	93.58	94.67	95.06	95.28	95.43	94.01
BS-A: VT	87.56	92.73	94.33	94.95	95.31	95.51	93.40
BS-A: VRP+TRP+VT	91.65	94.44	95.69	96.10	96.24	96.36	95.08
BS-B: VRP+TRP+VT	91.52	94.65	95.81	96.38	96.73	97.03	95.35
BS-C: VRP+TRP+VT	92.26	94.62	95.83	96.24	96.42	96.47	95.31
BS-A/B/C: VRP+TRP+VT	96.08	97.74	98.19	98.26	98.30	98.34	97.82

Architecture comparison on unseen subtypes (%)

Scheme	3 dB	8 dB	13 dB	18 dB	23 dB	No noise	Avg.
Proposed	96.08	97.74	98.19	98.26	98.30	98.34	97.82
Swin-B + Mean Fusion	94.60	96.90	97.82	98.07	98.23	98.38	97.33
ConvNeXt-B + Mean Fusion	93.90	95.73	96.65	97.14	97.45	97.69	96.43
ViT-B/16 + Mean Fusion	93.65	95.39	95.86	96.18	96.26	96.41	95.62

Conclusion

Key Contributions

Formulate the motion processes, echo channels, and echo signals for UAVs, birds, vehicles and pedestrians under multi-BS collaborative sensing scenario.
Propose a multi-scale time-frequency feature extraction algorithm for ISAC system, which extracts the VRP-TF, TRP-TF, and VT-TF of LAT as complementary multi-scale features.
Design a multi-BS and multi-scale feature fusion enabled LAT recognition network with Swin Transformer, which employs the visualized images of multi-scale feature to jointly recognize the target through deep feature extraction, intra-BS feature interaction, inter-BS feature interaction, and target recognition output.
Generate a massive echo signal dataset comprising 1,440,000 samples for LAT recognition within ISAC network.

Key Experimental Results

99.88% average accuracy on seen subtypes

97.82% average accuracy on unseen subtypes

96.08% accuracy at 3 dB on unseen subtypes

Experimental results demonstrate that multi-scale feature fusion can enhance the accuracy of LAT recognition, and multi-base station fusion can further boost the accuracy of LAT recognition.