Competitors

Team name: THWS

Team members: Max Werner, Sebastian Baader, Leander Knüttel, Frank Deinzer

Affiliation: Technical University of Applied Sciences Wuerzburg-Schweinfurt

Description:
This technical description gives a short overview of the indoor localization system developed at the Technical University of Applied Sciences Würzburg-Schweinfurt.
Based on a CONDENSATION particle filter, the localization system is highly modular with respect to the sensors it is able to use.
Models exist for Bluetooth low energy beacons, Wi-Fi, and others.
Combined with a step and turn detection, it can also handle scenarios of total sensor outages.
All calculations are performed in real-time on a commercial smartphone even when using a high number of samples for approximation.

Team name: OmniLoc

Team members: Xing Fang, Xile Gao, Taiyu Li

Affiliation: Autonavi

Description:
Our system combines an offline ground-truth pipeline with an online fusion framework.
Survey trajectories are generated by anchoring smartphone IMU data with control points in a factor-graph optimization, yielding accurate labels for magnetic and Wi-Fi/BLE radio maps.
At runtime, three modules run in parallel: a neural-network PDR (arbitrary phone posture), magnetic-field matching, and Wi-Fi/BLE fingerprint likelihood matching.
A Bayesian filter fuses them with barometric and road-network constraints.
Evaluation criteria: success rate, building/floor accuracy, and P50/P75/P90 horizontal error .
In the Sands Resorts Macao deployment (4 malls, 14 floors, 67k Wi-Fi APs, 2.6k beacons): 99.92% success, 99.09% building, 98.69% floor accuracy, P50 6.0 m, P90 18.6 m.

Team name: Trailblazer

Team members: Lingxiang Zheng, Bojun Chen, Xin Chen, Qinghong Zhu, Kaipeng Hong, Liqinzhen Liao, Wenying Cao, Liqin Wu

Affiliation: Xiamen University,XMU ATR laboratory

Description:
Our system is a confidence-aware hybrid factor-graph solution for indoor positioning.
A neural network converts synchronized IMU windows into relative position, velocity, orientation, and covariance constraints between keyframes.
WiFi/BLE positions, magnetic or AHRS heading, stride, zero-velocity, and barometric height are represented as probabilistic factors and weighted by their uncertainties.
The estimate minimizes robust Mahalanobis residuals over continuous navigation states, while floor labels are inferred from transition costs, and height-floor consistency.
Fixed-lag optimization provides online results.
The design prioritizes accuracy, continuity, outlier resistance, and calibrated uncertainty.

Team name: THWS

Team members: Max Werner, Sebastian Baader, Markus Bullmann, Leander Knüttel, Frank Deinzer

Affiliation: Technical University of Applied Sciences Wuerzburg-Schweinfurt

Description:
This technical description gives a short overview of the indoor localization system developed at the Technical University of Applied Sciences Würzburg-Schweinfurt.
Based on a CONDENSATION particle filter, the localization system is highly modular with respect to the sensors it is able to use.
Models exist for Bluetooth low energy beacons, Wi-Fi, and others.
Combined with a step and turn detection, it can also handle scenarios of total sensor outages.
All calculations are performed in real-time on a commercial smartphone even when using a high number of samples for approximation.

Team name: BUAA-IPNav

Team members: Jinkun LI, Zongbin REN, Yang MENG, Xiaohan YIN, Hanye TIAN, Ruofan ZHAO, Dayu YAN

Affiliation: Hangzhou International Innovative Institute of Beihang Univeristy, Beihang Univeristy

Description:
Our system achieves accurate, robust, and continuous 3D indoor positioning through adaptive multi-source fusion.
Deep-learning-based motion-mode recognition supports pedestrian dead reckoning (PDR), which serves as the inertial trajectory baseline.
Barometric measurements estimate vertical displacement and floor changes, while Wi-Fi/Bluetooth radio-map matching and geomagnetic sequence matching periodically correct accumulated drift.
Meanwhile, floor-plan contour matching constrains the trajectory to feasible areas, improving both positioning accuracy and floor-level reliability in complex indoor environments.

Team name: NUAA-FCRI

Team members: Liu Jiaming, Qi Tengfei, Wang Xinru, Wang Minkang, Cao Jian, Li Haoyang, Zhou Zhonghui, Liu Aoyun, Ren Yuanhui, Changhui Jiang

Affiliation: Flight Control Research Institute, College of Automation Engineering, Nanjing University of Aeronautics and Astronautics

Description:
We propose a causal, online smartphone positioning system for Track 3 that fuses pedestrian dead reckoning with WiFi/BLE fingerprinting, barometric floor estimation and GNSS in a factor graph solved with GTSAM, assisted by learning-based modules.
Phone inertial data near 50 Hz, RSS scans and pressure readings from the EvaalAPI are processed in 0.5 s windows.
Step events, stride length and heading come from the inertial and magnetic data, with heading initialised during the organiser calibration phase.
They form PDR, fingerprint, floor and GNSS factors.
Offline-trained neural networks predict adaptive factor weights and detect corrupted RSS scans and magnetic disturbances, so unreliable measurements are down-weighted rather than discarded.
Estimates are submitted at 2 Hz as longitude, latitude and floor (1-6) with 12 decimals; no ground truth is used during scoring.

Team name: OmniLoc

Team members: Xing Fang, Xile Gao, Taiyu Li

Affiliation: Autonavi

Description:
Our Track 3 system is a fully online, causal fusion framework for smartphone indoor positioning.
Magnetic and Wi-Fi/BLE radio maps are built offline from the organizers' labeled testing trials.
At run time three modules run in parallel: a neural-network PDR (arbitrary phone posture); magnetic-field matching using PDR heading for device-to-world transformation; and Wi-Fi/BLE fingerprint likelihood matching with missing-evidence penalties.
A Bayesian filter fuses them with barometric.
The system runs via EvaalAPI online mode: data read once, in sequence, each estimate timestamped at the latest sensor reading.
Criteria: building/floor accuracy, P50/P75/P90 error. Reference: a 4-mall, 14-floor resort complex — 99.09% building, 98.69% floor accuracy, P50 6.0 m, P90 18.6 m.

Team name: Multimodal Navigation Group

Team members: Sai Thejeshwar Sharma, Aicha Karite, Philipp Hager, Abhay Joshi, Susanna Kaiser, Martin Schmidhammer, Christian Gentner

Affiliation: German Aerospace Center (DLR)

Description:
The system estimates the position and floor of a person walking through a multi-storey building from the sensors of the smartphone they carry.
It runs in the off-site online mode of Track 3, consuming the stream through the EvAAL API and returning a longitude, a latitude and an integer floor every 0.5 s.
Three criteria governed the design.
Each estimate uses only the records that arrived before the epoch being answered, so no output depends on later data.
The only reference position it consumes is the starting position supplied by the organizers, and rows carrying reference positions are removed before anything downstream can see them.
Because the two recording devices report different absolute received signal strength for the same signal, no method relies on absolute RSS being comparable between phones.
The estimator combines pedestrian dead reckoning, barometric floor tracking, radio.

Team name: XTU-NaviTrust

Team members: 随嘉豪 (Sui Jiahao), 陈润芝, 徐莹, 陈基君, 杨锦豪, 李张阳, 黄 鲜, 秦令剑, 朱袁炼, 黄璐 (Huang Lu)

Affiliation: Xiangtan University – Hunan Provincial Key Laboratory of Intelligent and Trusted Navigation and Positioning (Xiangtan University, Hunan, China)

Description:
The competing system is a foot-mounted pedestrian navigation system that fuses GNSS, IMU, magnetometer and barometric pressure measurements within an error-state extended Kalman filter (ESKF).
High-rate IMU data provide continuous strapdown propagation, while causal zero-velocity detection triggers ZUPT during foot-stance phases.
GNSS position and velocity provide absolute corrections when reliable.
Calibrated magnetic measurements are used opportunistically for yaw correction only when magnetic-field consistency tests are satisfied.
Barometric pressure provides relative height and floor transition constraints.
Sensor biases and model parameters are calibrated using the public testing trial and supplementary static/calibration data.
The system operates causally without future-data smoothing.

Team name: NUAA-FCRI

Team members: Liu Jiaming, Qi Tengfei, Wang Minkang, Wang Xinru, Li Haoyang, Cao Jian, Ren Yuanhui, Liu Aoyun, Zhou Zhonghui, Changhui Jiang

Affiliation: Flight Control Research Institute, College of Automation Engineering, Nanjing University of Aeronautics and Astronautics

Description:
We propose a causal, map-free pedestrian navigation system for Track 4 that integrates foot-mounted PDR with GNSS in a factor graph, assisted by learning-based modules.
Raw ULISS streams (IMU near 200 Hz, magnetometer and barometer at lower rates, GNSS PVT near 1 Hz) are processed in 0.5 s windows.
Step events, stride length and heading come from the foot-mounted inertial and magnetic data, with heading initialized from the organiser key points.
They form PDR, GNSS position and barometric floor factors; the graph is solved with GTSAM in a sliding-window form to keep the solution causal.
Offline-trained neural networks predict adaptive factor weights and detect GNSS faults and magnetic disturbances, so unreliable measurements are down-weighted rather than discarded.
No map, pre-computed trajectory or ground truth is used, and estimates are submitted at 2 Hz through the EvaalAPI.

Team name: Multimodal Navigation Group

Team members: Philipp Hager, Sai Thejeshwar Sharma, Abhay Joshi, Susanna Kaiser, Aicha Karite, Martin Schmidhammer, Christian Gentner

Affiliation: German Aerospace Center (DLR)

Description:
Our system tracks a pedestrian from the ULISS foot-mounted IMU alone, with no maps and no radio infrastructure.
A 15-state error-state Kalman filter propagates the inertial solution in a local East-North-Up frame.
Stance phases, detected from smoothed gyroscope magnitude, trigger zero-velocity updates that bound the drift, and BMP280 pressure gives the floor by hypsometric levelling.
GNSS never enters the filter. Surviving fixes are fitted to the dead-reckoned track by a single-angle Procrustes rotation, setting only the orientation of the output frame while the position stays anchored on the first surveyed key point.
The client runs online in 0.5 s windows through the EvaalAPI, committing each estimate before it reads the next.
On the 2026 practice trial it scores 4.64 m third-quartile error with every floor correct.

Team name: OmniLoc

Team members: Shaun Fang

Affiliation: Autonavi

Description:
Traditional positioning algorithms face critical limitations in complex environments such as urban canyons, tunnels, and indoor spaces.
To address these challenges, Amap proposed the industry's first end-to-end multi-source fusion positioning model.
Built on a Transformer architecture, it fuses GNSS, IMU, network location, barometer, and road network features via cross-attention weighting, jointly outputting position, velocity, matched road, and motion mode.
The model significantly outperforms traditional algorithms: reducing urban canyon errors to 3–8 m, maintaining sub-15 m accuracy through 60 s of signal outage, requiring no warm-up phase, generalizing across devices, and seamlessly supporting walking, driving, and cycling.
It has been deployed in the Amap App, generating hundreds of billions of positioning points every day.

Team name: SmartPositioning

Team members: Haoquan Sun, Yue Yu, Shiyu Bai, Mengyu Ding, Jingxian Wang

Affiliation: Shanghai University, The Hong Kong Polytechnic University

Description:
This system delivers 3D pedestrian and vehicle positioning in urban areas using a smartphone.
It fuses GNSS, IMU and barometer measurements in a loosely coupled Extended Kalman Filter.
For vehicles, a CNN GRU DeepOdo network estimates forward velocity without wheel sensors or CAN bus access.
Non holonomic, zero velocity and zero integrated heading rate constraints reduce inertial drift, while the barometer supports altitude and floor changes.

Team name: BUPT_YuBoTongXing

Team members: Yiran Zhang, Kaile Yu, Zhixing Ren, Yancong Peng, Weizheng Ren

Affiliation: Beijing University of Posts and Telecommunications

Description:
Our system estimates 2D UE positions from 5G channel impulse responses collected by distributed base stations.
It preprocesses complex CIR measurements into robust log-magnitude and delay-domain features, combined with ToA, antenna masks, and timestamps.
To address sparse references, the method uses self-supervised channel representation learning and channel charting on unlabeled data, followed by lightweight supervised regression using the available labeled samples.
Missing antennas and unreliable CIR/ToA measurements are handled with masks, outlier detection, and robust filtering.
Temporal consistency and speed constraints are applied within each recording segment.
The system outputs 2D position estimates and is optimized for low Euclidean localization error, especially the third-quartile error.

Team name: CUMT-HC

Team members: Teng Wang, Guoliang Chen, Jinyu Dai, Jiaqi Dong, Meng Sun, Jingxue Bi, Zitao Lin, Wancong Liu, Ruimin Ma, Dongjun Lv

Affiliation: School of Environment and Spatial Informatics, China University of Mining and Technology

Description:
Our system provides an indoor localization framework utilizing a self-supervised Masked Autoencoder and temporal smoothing to address sparse reference challenges.
First, we extract physically grounded channel features to mitigate signal degradation and synchronization issues.
We then pretrain the autoencoder on unlabelled data to reconstruct masked signals, distilling complex multipath phenomena into robust representations.
In the regression stage, a frozen encoder feeds a shallow Multi-Layer Perceptron trained exclusively on the 50 labeled reference samples.
Finally, dynamic temporal smoothing filters high-frequency multipath jitter to ensure kinematically realistic trajectory estimation.

Team name: Team POSTECH

Team members: Hanjun Park

Affiliation: Pohang University of Science and Technology (POSTECH)

Description:
Euclidean distance between estimated and true results is the error metric, and CE75 is used as a performance metric.
We use EvaalAPI for the competition.
One day prior to the competition, an AES-GCM encrypted version of the competition test dataset will be uploaded to the owncloud.
During the competition, participants can retrieve the decryption key from the EvaalAPI via the nextdata endpoint.
Once decrypted, participants run their localization algorithm on the test dataset and upload the predicted positions back through the EvaalAPI.
We will have 300 seconds for the competition.

Team name: Skygooo

Team members: Tian Sun

Affiliation: Aerospace Information Technology University

Description:
Our system performs offline 2D indoor localization using measurements from eight radio units and only 50 labeled training positions.
It constructs relative time-of-arrival fingerprints by removing common timing offsets, rejecting outliers, and applying temporal filtering within each recording.
Nonlinear slow feature analysis learns compact representations at multiple temporal scales from unlabeled observations.
Kernel ridge regression then maps these representations to physical coordinates using the sparse position labels.
Feature configurations, regression parameters, and ensemble size are selected through training-label leave-one-out evaluation, using a combination of mean Euclidean error and 75th-percentile error (CEP75).
Selected predictions are averaged and refined using a constant-velocity Kalman model, Rauch–Tung–Striebel smoothing, and speed constraints.

Team name: MIUN Pulse

Team members: Syed Muhammad Jameel, Hamed Talebian, Aamir Mahmood

Affiliation: Department of Computer and Electrical Engineering, Mid Sweden University

Description:
The proposed CIRFormer is a transformer-based 2D localization system for IPIN 2026 Track 7 using only 50 labeled references and about 259k unlabeled CIR samples.
Each valid RU CIR is RMS-normalized and split into 12-tap patches, and we add centred ToA, RU identity, patch position, RSS, and signal-quality features to the token representation.
Training has three stages: masked self-supervised reconstruction of CIR log-magnitude and adjacent-tap phase increments; label-free metric learning using temporal ordering, PDP/ToA similarity, hard negatives, and augmentation consistency; and 2D position learning using sparse labels, confidence-weighted TDoA pseudo-targets, and motion constraints.
Final predictions are temporally smoothed independently within each recording.