Organizations: Wenhui Feng, Muhan Weng, and Jianyu Liu are with the Electronic Information School, Wuhan University, Wuhan, Hubei 430072, China · Xiansheng Yang, Hao Zhang, and Yuan Zhuang are with the State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing (LIESMARS), Wuhan University, Wuhan, Hubei 430072, China · Liang Wang is with the Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100094, China
Smartphone indoor localization requires low-cost, robust positioning across scene transitions and phone-use variations. Pedestrian dead reckoning (PDR) enables infrastructure-free deployment but accumulates errors from heading drift, structural transitions, and carrying-mode changes. Existing methods often treat contextual cues as a single event stream or local correction signal, limiting their ability to address distinct localization errors. This paper presents a framework enhanced by hierarchical multi-sensor context recognition. A jointly data- and model-driven recognition model predicts environment, scene, and activity probabilities, which are converted into scene-event and action-context segments. Scene-event segments guide scene-constrained PDR correction through corridor heading constraints, landmark attraction, fixed-lag backtracking, and vertical-transition handling. Action-context segments drive action-aware RoNIN-based position compensation for carrying-mode-dependent drift. Experiments on cross-floor and multi-action trajectories against LiDAR-SLAM references show that the proposed framework outperforms PDR and representative learned inertial odometry baselines, demonstrating that hierarchical semantics provide reliable, localization-specific correction cues.
Figures & tables
Fig. 1: Overview of the key challenges and the proposed smartphone localization framework enhanced by hierarchical multi-sensor context recognition.
Method family
Context cue and use
Our Work
Context-aware PDR
Motion state, carrying mode, or map structure adjusts steps, headings, and floor transitions.
Scene events drive map-guided PDR correction.
Semantic-aided localization
Scene, activity, or landmark events constrain routes, topology, or graph factors.
Environment, scene, and activity cues are separated before localization.
Learned inertial odometry
IMU windows predict velocity or displacement in a learned trajectory stream.
TABLE I: Context Cues in Related Localization Methods
Fig. 2: Hierarchical multi-sensor context recognition model. The model extracts temporal and mask-aware statistical features for each sensor, aggregates cross-sensor relations with a fully connected graph-attention module, and decodes environment, scene, and activity probabilities with top-down semantic constraints.
Fig. 3: Auxiliary localization framework driven by recognition-derived semantic cues. Scene-event segments activate scene-constrained auxiliary localization for PDR, while action-context segments guide action-aware RoNIN-based learned position compensation.
Sensor
Recorded attributes
Sampling rate
Accelerometer
3-axis linear acceleration
100–200 Hz
Gyroscope
3-axis angular velocity
100–200 Hz
Magnetometer
3-axis magnetic field
100 Hz
AHRS
Pitch, roll, yaw, quaternion
100–200 Hz
Barometer
Pressure
20 Hz
Light
Illuminance
5–10 Hz
TABLE II: Smartphone Sensor Configuration Used in the Experiments
Parameter
Value
Semantic and elevator confidence threshold, τq
0.9
Minimum transition-event persistence, Nmin
2 consecutive windows
Corridor pull strength, λ
0.6
Corridor heading tolerance, ε
20∘
Maximum position correction, dmax
20.0 m
Correction regularization coefficient, β
0.005 (training)
TABLE III: Principal Parameters Used in the Localization Experiments
Model
Environment Accuracy
Scene Accuracy
Activity Accuracy
Joint Accuracy
Mean Macro-F1
Ours
97.54
95.24
98.06
94.35
96.83
TABLE IV: Recognition Performance of the Semantic Front End (%)
Method
Description
PDR
Step-wise dead reckoning baseline driven by pedestrian motion estimation and heading updates.
RoNIN
Robust neural inertial navigation baseline used as the main learned trajectory reference.
MEIO
Motion-aware inertial odometry baseline with adaptive computation design.
RIDI
Double-integration-based inertial baseline representing a classic learning-based IMU pipeline.
TLIO
Tight learned inertial odometry baseline for drift-aware trajectory estimation.
Our Work
Hierarchical semantic interface with scene-constrained PDR correction on Route A and action-aware RoNIN position compensation on Route B.
TABLE V: Compared Methods Used in the Experiments
Method
ATE (m)
RTE@10 m (m)
APE med (m)
Route A
PDR
17.705
12.298
7.974
TLIO
8.666
6.164
5.214
MEIO
5.537
5.293
7.593
RIDI
7.280
6.933
4.307
RoNIN
3.727
4.738
1.945
Our Work
2.699
3.765
1.524
TABLE VI: Overall Quantitative Comparison on Route A and Route B
Fig. 4: Route A trajectory comparison across floor transitions. (a) Initial trajectory segment on the second floor. (b) Trajectory segment after descending to the first floor. (c) Trajectory segment after returning to the second floor.
Fig. 5: Route B overall trajectory comparison under multiple phone-use switches.
Fig. 6: Overall error analysis. The upper row shows displacement-error CDFs, and the lower row shows APE time series. (a) Route A displacement-error CDF. (b) Route B displacement-error CDF. (c) Route A APE time-series comparison. (d) Route B APE time-series comparison.
Fig. 7: Localized gain analysis around semantic triggers. The upper row shows APE curves in [−15,25] s windows, and the lower row shows APE distributions. (a) Route A key-point APE evolution. (b) Route B action-switch APE evolution. (c) Route A key-point APE distribution. (d) Route B action-switch APE distribution.
TABLE VII: Ablation results for Route A scene-constrained components and Route B action-aware components.