PanoAir: A Panoramic Visual-Inertial SLAM with Cross-Time Real-World UAV Dataset

Yiyang Wu; Xiaohu Zhang; Yanjin Du; Tongsu Zhang; Chujun Li; Siyang Chen; Guoyi Zhang; Xiangpeng Xu

cs.ROApr 1, 2026

PanoAir: A Panoramic Visual-Inertial SLAM with Cross-Time Real-World UAV Dataset

Authors: Yiyang Wu, Xiaohu Zhang, Yanjin Du, Tongsu Zhang, Chujun Li, Siyang Chen, Guoyi Zhang, Xiangpeng Xu

Organizations: School of Aeronautics and Astronautics, Sun Yat-sen University, Guangzhou 510275, China.

Abstract

Accurate pose estimation is fundamental for unmanned aerial vehicle (UAV) applications, where Visual-Inertial SLAM (VI-SLAM) provides a cost-effective solution for localization and mapping. However, existing VI-SLAM methods mainly rely on sensors with limited fields of view (FoV), which can lead to drift and even failure in complex UAV scenarios. Although panoramic cameras provide omnidirectional perception to improve robustness, panoramic VI-SLAM and corresponding real-world datasets for UAVs remain underexplored. To address this limitation, we first construct a real-world panoramic visual-inertial dataset covering diverse flight conditions, including varying illumination, altitudes, trajectory lengths, and motion dynamics. To achieve accurate and robust pose estimation under such challenging UAV scenarios, we propose a panoramic VI-SLAM framework that exploits the omnidirectional FoV via the proposed panoramic feature extraction and panoramic loop closure, enhancing feature constraints and ensuring global consistency. Extensive experiments on both the proposed dataset and public benchmarks demonstrate that our method achieves superior accuracy, robustness, and consistency, completing all 17 sequences without tracking failure and reducing the ATE by up to 47.1% compared with the best-performing baseline. Moreover, deployment on an embedded platform validates its practical applicability, achieving comparable computational efficiency to PC implementations. The source code and dataset are publicly available at https://github.com/YYWumm/PanoAir

Figures & tables

Dataset	Publication	Setting	Panoramic	IMU	Real-world	Ground-truth
OpenVSLAM [ 15 ]	MM’19	Handheld	✓	✗	✓	✗
Kitti-360 [ 25 ]	TPAMI’22	Ground Vehicle	✗	✓	✓	SFM & GPS
360VO [ 17 ]	ICRA’22	Virtual scene	✓	✗	✗	Synthetic
360VIO [ 18 ]	TII’23	Ground Vehicle & Handheld	✓	✓	✓	MoCap & RTK
Pair360 [ 23 ]	RAL’24	Ground Vehicle	✓	✓ –	✓	GPS
TartanAir-V2 [ 26 ]	IROS’25	Virtual scene	✓	✓	✗	Synthetic

TABLE I: Comparison of existing benchmarks for panoramic visual and visual-inertial odometry/slam

Fig. 1: Overview of the proposed dataset. A panoramic camera is mounted on a UAV to collect panoramic-IMU data, with ground-truth trajectories recorded using an RTK ground station (a). The dataset covers diverse cross-time scenarios with varying lighting conditions, altitudes, flight speeds, and maneuvering characteristics, as summarized in (b), with representative scenes shown in (c). Further details are provided in Section III .

Sequence	Condition	Avg. Speed (m/s)	Length (m)	Duration (s)
Seq 1	Cloudy	3.0	257	87
Seq 2		5.5	822	150
Seq 3		7.3	1168	159
Seq 4		8.0	1262	158
Seq 5		10.0	1602	160
Seq 6	Sunny	4.0	803	200

TABLE II: Statistics of different sequences of the proposed dataset

Sensor	Parameter	Rate
Camera	Rolling Shutter Panoramic: 1280x640; Front fisheye: 640x640; Pinhole: 980x640	30 Hz
IMU	6-axis Accelerometer noise density: $5.88\text{\times}{10}^{-4}\text{\,}\mathrm{m}\text{\,}{\mathrm{s}}^{-2}\text{\,}{\mathrm{}}^{-1}$ Accelerometer random walk: $1.84\text{\times}{10}^{-4}\text{\,}\mathrm{m}\text{\,}{\mathrm{s}}^{-3}\text{\,}{\mathrm{}}^{-1}$ Gyroscope noise density: $8.29\text{\times}{10}^{-5}\text{\,}\mathrm{rad}\text{\,}{\mathrm{s}}^{-1}\text{\,}{\mathrm{}}^{-1}$ Gyroscope random walk: $7.85\text{\times}{10}^{-5}\text{\,}\mathrm{rad}\text{\,}{\mathrm{s}}^{-2}\text{\,}{\mathrm{}}^{-1}$	1000 Hz
RTK	Horizontal: 1 cm+ 1 ppm(RMS); Vertical: 2 cm+ 1 ppm(RMS)	5 Hz

TABLE III: Statistics of Sensors of the proposed dataset

Fig. 2: Time synchronization between the panoramic camera and the UAV platform.

Fig. 3: Overview of the proposed panoramic visual-inertial SLAM framework. The system takes panoramic images and IMU measurements as inputs, and achieves robust, accurate, and globally consistent results through panoramic feature extraction (Section IV-A ), as well as panoramic optimization and panoramic loop closure (Section IV-B ). Experimental results are presented in Section V .

Fig. 4: Hybrid feature tracking on ERP images. Red dots denote learning-based features (e.g., SuperPoint), green dots denote hand-crafted features (e.g., ORB).

Method	Settings		Cloudy					Sunny				Night								Avg.	Succ.
Method	Sensor	LC	Seq1	Seq2	Seq3	Seq4	Seq5	Seq6	Seq7	Seq8	Seq9	Seq10	Seq11	Seq12	Seq13	Seq14	Seq15	Seq16	Seq17	Avg.	Succ.
ORB-SLAM3 [ 1 ]	FI	✓	0.800	0.596	✗	3.969	✗	0.403	0.293	0.633	1.874	3.503	0.907	1.019	0.868	4.539	3.030	✗	1.637	✗	83%
VINS-mono [ 2 ]	FI	✓	0.555	1.085	1.889	2.835	2.816	0.359	0.344	1.153	3.417	✗	✗	✗	✗	✗	✗	✗	10.930	✗	59%
Droid-SLAM * [ 47 ]	P	—	0.243	0.369	0.435	✗	1.211	0.423	0.253	11.387	1.046	0.519	3.970	0.675	0.373	✗	✗	✗	15.425	✗	76%
OpenVSLAM * [ 15 ]	O	✓	0.370	0.535	0.496	0.703	0.758	0.722	0.449	0.544	1.294	0.478	0.864	✗	0.593	0.382	0.755	0.799	1.397	✗	94%
360DVO * [ 28 ]	O	✗	0.388	1.199	1.119	1.677	3.677	1.259	0.247	0.893	3.927	1.044	0.978	0.696	0.298	0.828	0.984	2.582	1.963	1.398	100%

TABLE IV: Tracking accuracy on our proposed panoramic UAV benchmark (ATE,in meters)

Method	Sequence	Successful Runs	ATE RMSE (m)
ORB-SLAM3 [ 1 ]	Seq3	2/5	$4.129\pm 0.122$
	Seq5	1/5	$4.383$
	Seq16	2/5	$1.463\pm 0.283$
VINS-Mono [ 2 ]	Seq10	0/5	–
	Seq11	0/5	–
	Seq12	0/5	–

TABLE V: Repeated-run statistics for failed cases in Table IV

Methods	Settings		Seq 1 (Easy)		Seq 2 (Medium)		Seq 3 (Hard)		Avg.
Methods	Sensor	LC	ATE	RPE	ATE	RPE	ATE	RPE	ATE	RPE
ORB-SLAM3 [ 1 ] †	FI	✓	0.186	0.108	0.256	0.068	0.186	0.086	0.209	0.087
VINS-mono [ 2 ] †	FI	✓	1.406	0.082	2.960	0.242	✗	✗	✗	✗
OpenVSLAM * [ 15 ]	O	✗	0.149	0.049	0.058	0.068	✗	✗	✗	✗
360DVO * [ 28 ]	O	✗	0.184	0.049	0.217	0.067	0.520	0.089	0.307	0.068
360-VIO [ 18 ] †	OI	✗	0.456	0.058	0.390	0.088	0.632	0.122	0.493	0.089

TABLE VI: Tracking accuracy on 360-VIO benchmark (ATE and RPE, in meters)

Fig. 5: Qualitative comparison results. Top: comparison of different methods on the proposed dataset, where our method produces estimates closer to the ground truth and demonstrates stronger robustness. Bottom: results on an additional large-scale handheld sequence. Other methods suffer from varying degrees of drift, while our method achieves lower drift and, by leveraging panoramic loop closure, successfully aligns with the starting position, resulting in a globally consistent estimation.

End position (x,y,z in meters)	Handheld indoor 330 m	Handheld outdoor 5.5 km
Ours(Without Loop)	-1.52, -0.53, -0.74	-9.92, -49.37, -28.43
Ours(With Loop)	-0.01, 0.00, 0.01	-1.39, -0.08, 0.36

TABLE VII: End positions of two handheld sequences with aligned start and end positions.

Setting	Cloudy	Sunny	Night	Indoor	Total count
With positive-depth check	18	20	16	10	64
Without positive-depth check	21	23	17	11	72

TABLE VIII: Detected loop closure counts with and without positive-depth check

Fig. 6: Box plots of ATE over five runs on the PC with 10 Hz and 30 Hz frame inputs, and on the Orin NX with 10 Hz frame input. The error variations remain stable across different platforms and input frequencies, demonstrating the feasibility of onboard deployment.

Method	ATE(m) $\downarrow$
Method	Cloudy	Sunny	Night	Indoor	Avg.
w/o learning-based feature	1.095	0.558	0.930	✗	✗
w/o hand-crafted feature	1.566	0.553	✗	0.574	✗
w/o distortion weight	1.342	0.556	0.789	0.150	0.709
w/o panoramic loop closure	2.170	1.087	1.517	0.145	1.223
w/o omnidirectional FOV	1.482	0.599	✗	0.322	✗

TABLE IX: Ablation results of the proposed module. Averages computed over all sequences in different scenarios.

Runtime(ms)	Cloudy(Seq5)			Sunny(Seq9)			Night(Seq17)			Avg.
Runtime(ms)	PC 1	PC 2	OrinNX 3	PC 1	PC 2	OrinNX 3	PC 1	PC 2	OrinNX 3	PC 1	PC 2	OrinNX 3
HybridFeature Extraction	14.0	14.1	76.5	13.9	13.6	74.3	12.4	12.7	65.6	13.4	13.5	72.1
Local Optimization	9.0	6.9	32.8	9.2	7.0	33.1	7.9	6.4	29.7	8.7	6.8	31.9
Total Tracking	25.6	23.1	118.7	26.1	23.3	119.8	22.9	20.8	104.0	24.9(40fps)	22.4(45fps)	114.2(9fps)

TABLE X: Runtime statistics of different platforms evaluated on three long trajectories from the proposed dataset.

Visual-Inertial SLAM

Visual-inertial simultaneous localization and mapping (SLAM) combines camera and inertial data to estimate motion and build consistent maps.

See all 29 papers

Sep 28, 2026

ForVis: An In-Field Dataset and Benchmark for VIO Using Under-Canopy UAV Flights in Forests

Visual-inertial Simultaneous Localization and Mapping (VI-SLAM) for UAVs remains difficult to evaluate in real forest environments, where motion, illumination changes, repetitive vegetation, and vibration can all affect estimation. We present ForVis, an in-field dataset and benchmark for evaluating VI-SLAM during UAV flight in forest environments. The dataset contains twelve flights across open meadow, above-canopy, and under-canopy conditions in each environment. In total, it provides 563.8s of flight over 1096.8m of trajectory, recorded simultaneously with an Intel RealSense D435i and an OAK-D Pro Wide together with inertial and flight-controller data. We benchmark seven open-source VI-SLAM systems over 504 runs. The results show that sensor choice has a larger effect on trajectory error than the spread between algorithms: all seven methods achieve lower median error on the OAK-D Pro than on the D435i. ForVis is intended to support evaluation of speed, accuracy and robustness for VI-SLAM in challenging forest flight.

Arman Kiani, Masoud Ataei, Elvis Gyaase +4

Department of Electrical and Computer Engineering, University of Maine, Orono, ME 04469, USA · School of Forest Resources, University of Maine, Orono, ME 04469, USA

UAV Navigation Visual-Inertial SLAM Visual-Inertial Odometry

Ask PDF

May 5, 2026

Robust Visual SLAM for UAV Navigation in GPS-Denied and Degraded Environments: A Multi-Paradigm Evaluation and Deployment Study

Reliable localization in GPS-denied, visually degraded environments is critical for autonomous UAV opera- tions. This paper presents a systematic comparative evaluation of five V-SLAM systems ORB-SLAM3, DPVO, DROID-SLAM, DUSt3R, and MASt3R spanning classical, deep learning, recurrent, and Vision Transformer (ViT) paradigms. Experiments are conducted on curated sequences from four public benchmarks (TUM RGB-D, EuRoC MAV, UMA-VI, SubT-MRS) and a custom monocular indoor dataset under five controlled degradation conditions (normal, low light, dust haze, motion blur, and combined), with sub-millimeter Vicon ground truth. Results show that ORB-SLAM3 fails critically under severe degradation (62.4% overall TSR; 0% under dense haze), while learning-based methods remain robust: MASt3R achieves the lowest degraded ATE (0.027 m) and DUSt3R the highest tracking success (96.5%). DPVO offers the best efficiency robustness trade-off (18.6 FPS, 3.1 GB GPU memory, 86.1% TSR), making it the preferred choice for memory-constrained embedded platforms. Embedded deployment analysis across NVIDIA Jetson platforms provides actionable guidelines for SLAM selection under SWaP-constrained UAV scenarios.

Prasoon Kumar, Akshay Deepak, Sandeep Kumar

National Institute of Technology Patna, Bihar, India · Central Research Laboratory, Bharat Electronics Ltd., Ghaziabad, Uttar Pradesh, India

UAV Navigation GNSS-Denied Localization Visual SLAM

Ask PDF

Oct 3, 2026

Synthetic-to-Real ViT-Based Pose Estimation of a Noncooperative UAV

Remote pose estimation of noncooperative Unmanned Aerial Vehicles (UAVs) from imagery is critical, as they cannot be influenced or instrumented in advance. Deep-learning-based approaches offer a promising solution; however, their development is constrained by the cost and difficulty of acquiring large-scale real-world datasets with accurate pose labels. Synthetic imagery provides an alternative, but models trained on synthetic data must overcome the synthetic-to-real domain gap to generalize to real-world imagery. This work investigates the inherent synthetic-to-real generalization capability of a Vision Transformer (ViT)-based model for monocular UAV pose estimation. The proposed approach employs a self-supervised DINOv2 backbone and is trained exclusively on labeled synthetic imagery while being evaluated on labeled real-world imagery. Pose ambiguity-aware strategies are incorporated during training and inference to address ambiguities arising from the projection of a three-dimensional target onto a two-dimensional image plane and from target symmetries. An

α

β

filter is further integrated during inference to improve pose estimations. To assess the model under operational requirements, it is evaluated in terms of Mean Angular Error (MAE) and inference time, both before and after filtering, using a real-world dataset containing 77,077 labeled UAV images. Before filtering, the model achieves an MAE of

19.18^{\circ}

and an inference time of

13.25

ms, whereas after filtering, these values are

8.74^{\circ}

and

13.42

ms, respectively.

Krishnanujam Srinivas, Hanish Acharla, Brij Agrawal +1

Spacecraft Research and Design Center, Department of Mechanical and Aerospace Engineering, Naval Postgraduate School, Monterey.

Object Pose Estimation Synthetic-to-Real Domain Adaptation Vision Transformer

Ask PDF

Abstract

Visual-Inertial SLAM

Explore similar work

ForVis: An In-Field Dataset and Benchmark for VIO Using Under-Canopy UAV Flights in Forests

Robust Visual SLAM for UAV Navigation in GPS-Denied and Degraded Environments: A Multi-Paradigm Evaluation and Deployment Study

Synthetic-to-Real ViT-Based Pose Estimation of a Noncooperative UAV