Accurate Pedestrian Tracking in Urban Canyons: A Multi-Modal Fusion Approach

The contribution describes a pedestrian navigation approach designed to improve localization accuracy in urban environments where GNSS performance is degraded, a problem that is especially critical for blind or low-vision users who depend on precise guidance such as identifying the correct side of a street. To address GNSS limitations and the impracticality of camera-based visual positioning, the work proposes a particle filter based fusion of GNSS and inertial data that incorporates spatial priors from maps, such as impassable buildings and unlikely walking areas, functioning as a probabilistic form of map matching. Inertial localization is provided by the RoNIN machine learning method, and fusion with GNSS is achieved by weighting particles based on their consistency with GNSS estimates and uncertainty. The system was evaluated on six challenging walking routes in downtown San Francisco using three metrics related to sidewalk correctness and localization error. Results show that the fused approach (GNSS+RoNIN+PF) significantly outperforms GNSS only localization on most metrics, while inertial-only localization with particle filtering also surpasses GNSS alone for critical measures such as sidewalk assignment and across street error.

💡 Research Summary

**
The paper addresses the well‑known problem of degraded GNSS performance in dense urban “canyon” environments, where tall buildings block line‑of‑sight to satellites and cause multipath and attenuation. This degradation is especially critical for blind or low‑vision pedestrians who rely on precise navigation cues such as identifying the correct side of a street. Visual positioning services that use the phone camera are impractical when the device is kept in a pocket, so the authors propose a purely sensor‑based solution that fuses GNSS with inertial navigation while exploiting map priors.

The core of the system consists of three components: (1) an inertial navigation module based on RoNIN (Robust Neural Inertial Navigation), a deep‑learning model that directly predicts 2‑D velocity from 200 Hz IMU data in a fixed world reference frame; (2) a particle filter that maintains 500 hypotheses (particles) about the user’s position, orientation drift, and weight; and (3) spatial priors derived from GIS data that label map regions as “impenetrable” (buildings), “freely traversable” (sidewalks and crosswalks), or “street” (road surface outside marked crosswalks). Particles located inside buildings receive zero weight and are eliminated, particles on sidewalks receive weight 1, and particles on street surfaces receive an intermediate weight (empirically set to 0.4) to reflect the lower likelihood of jaywalking.

The particle filter propagates each particle using the RoNIN velocity vector, rotated by the particle’s orientation‑drift estimate, and adds Gaussian noise to both position and drift. After propagation, particle weights are updated based on the map prior and, when GNSS uncertainty is low, by multiplying the weight with a bivariate Gaussian centered on the GNSS fix. The Gaussian’s standard deviations are proportional to the GNSS horizontal accuracy reported by iOS. This step pulls particles toward the GNSS measurement when it is reliable, while allowing the inertial estimate to dominate when GNSS is noisy or unavailable. The final user position is the weighted average of all particles.

The authors collected data on six routes in downtown San Francisco (lengths 309 m–644 m) that traverse areas with severe GNSS blockage, including alleyways and the covered Salesforce Transit Center. Ground‑truth positions were recorded at each detected footstep. Three configurations were evaluated: GNSS‑only, RoNIN + particle filter (RoNIN+PF), and the full fusion (GNSS + RoNIN+PF). Performance was measured with three metrics: (i) Correct Sidewalk Assignment – the proportion of time the estimated location lies on the correct sidewalk within the correct city block; (ii) Euclidean error – the straight‑line distance between estimate and ground truth; and (iii) Along‑ and Across‑Street error – the decomposition of Euclidean error into components parallel and perpendicular to the nearest street direction.

Results show that the fused approach (GNSS + RoNIN+PF) consistently outperforms GNSS‑only on all metrics except Along‑Street error, where differences are marginal. Notably, Across‑Street error—a common failure mode in urban canyons—is dramatically reduced, and sidewalk assignment accuracy rises above 90 % on most routes. Interestingly, the inertial‑only solution with particle filtering also surpasses GNSS‑only for sidewalk assignment and Across‑Street error, demonstrating that map‑aware particle filtering can compensate for the lack of absolute GNSS fixes by enforcing realistic walking constraints.

Key contributions of the work are: (1) integration of a state‑of‑the‑art deep‑learning inertial navigation model with a particle‑filter framework; (2) incorporation of building and road priors as probabilistic map‑matching weights, which effectively eliminates physically impossible hypotheses; (3) a systematic experimental evaluation in a real‑world urban canyon setting that quantifies the complementary strengths of GNSS and inertial data. The study validates that a relatively simple weighting scheme—Gaussian scaling by GNSS uncertainty combined with map‑based particle weights—can handle non‑Gaussian, non‑linear error characteristics typical of pedestrian navigation in dense cities.

Future directions suggested include extending the map prior to three‑dimensional building models for shadow‑matching, integrating dynamic intent prediction (e.g., upcoming turns or crosswalk usage) via additional machine‑learning modules, and optimizing particle count and resampling frequency for energy‑efficient deployment on mobile devices. Overall, the paper demonstrates a practical, low‑cost solution that markedly improves pedestrian localization for vulnerable users in challenging urban environments.