PaperlayerView exact revision on arXiv

Source-linked AI summary

Think over Trajectories: Leveraging Video Generation to Reconstruct GPS Trajectories from Cellular Signaling

Ruixing Zhang, Hanzhang Jiang, Leilei Sun, Liangzhe Han, Jibin Wang, Weifeng Lv

arXiv:2603.26610v1cs.CVcs.AI

TL;DR

Sig2GPS addresses the limited spatial precision of cellular signaling by reconstructing fine-grained GPS trajectories. The paper uses map-visual video generation, fine-tunes on paired signaling-trajectory videos, and applies trajectory-aware verifiable-reward optimization. Experiments report gains over engineered and learning-based baselines, with additional evidence of scalability and cross-city transferability.

  • Problem

    Cellular signaling provides broad coverage but coarse spatial cues, creating a need to reconstruct high-precision GPS trajectories for mobility applications.

  • Method

    The paper renders signaling on maps, fine-tunes a video generator on paired signaling-trajectory videos, and applies Traj-GDPO with trajectory-aware verifiable rewards.

  • Results

    Experiments report consistent gains over learning-based baselines and a production-grade engineered pipeline, while next-GPS prediction indicates scalability and cross-city transferability.

  • Takeaways & Limitations

    Map-visual video generation offers a practical interface for generating and refining continuous, topology-consistent trajectories under map constraints.

  • Takeaways & Limitations

    The evaluation assumes evenly sampled GPS trajectories with a fixed sampling interval, and an industrial baseline remains a competitive multi-stage pipeline.

Abstract

from arXiv · show

Mobile devices continuously interact with cellular base stations, generating massive volumes of signaling records that provide broad coverage for understanding human mobility. However, such records offer only coarse location cues (e.g., serving-cell identifiers) and therefore limit their direct use in applications that require high-precision GPS trajectories. This paper studies the Sig2GPS problem: reconstructing GPS trajectories from cellular signaling. Inspired by domain experts often lay the signaling trace on the map and sketch the corresponding GPS route, unlike conventional solutions that rely on complex multi-stage engineering pipelines or regress coordinates, Sig2GPS is reframed as an image-to-video generation task that directly operates in the map-visual domain: signaling traces are rendered on a map, and a video generation model is trained to draw a continuous GPS path. To support this paradigm, a paired signaling-to-trajectory video dataset is constructed to fine-tune an open-source video model, and a trajectory-aware reinforcement learning-based optimization method is introduced to improve generation fidelity via rewards. Experiments on large-scale real-world datasets show substantial improvements over strong engineered and learning-based baselines, while additional results on next GPS prediction indicate scalability and cross-city transferability. Overall, these results suggest that map-visual video generation provides a practical interface for trajectory data mining by enabling direct generation and refinement of continuous paths under map constraints.

1 Introduction

Sig2GPS addresses the coarse spatial resolution of cellular signaling by reconstructing fine-grained GPS trajectories. Think Over Trajectory reframes this reconstruction as map-visual video generation, combining paired-video fine-tuning with trajectory-aware reinforcement learning.

  • Motivation: Cellular signaling records typically identify serving base stations rather than precise user locations, limiting high-precision mobility applications.Transforming signaling into fine-grained GPS trajectories could broaden signaling data’s applicability to mobility analytics and services.
  • Motivation: Existing signal-to-GPS systems use long, multi-stage pipelines involving noise mitigation, map matching, and route inference.Heterogeneous environments also require case-specific heuristics, producing complex codebases that depend heavily on expert knowledge.
  • Think Over Trajectory: Think Over Trajectory casts Sig2GPS as image-to-video generation, drawing a continuous GPS path from map-based signaling visualizations.The paradigm mirrors domain experts’ practice of overlaying signaling traces on maps and sketching plausible routes.
  • Think Over Trajectory: Unlike approaches that regress coordinates or directly encode trajectories, the framework operates in the map-visual domain to model continuous path drawing.The paper argues this provides a more faithful abstraction of human trajectory reasoning than discrete-coordinate outputs.
  • Contributions: The framework fine-tunes a video generator on paired signaling-trajectory videos and applies Traj-GDPO using rewards for distance error, heading, and branching.Evaluation uses 10,000 paired signal-GPS trajectories and reports gains over industrial and learning-based methods, with next-GPS results indicating scalability and cross-city transferability.

2 Related Work

Related work spans conventional and deep-learning trajectory modeling, visual trajectory mining, and flow-based video generation. The paper positions its contribution as generating continuous map-drawn paths rather than numeric coordinates, while using video models’ spatial reasoning and verifiable rewards.

  • Trajectory Mining: Trajectory mining has progressed from Markov chains to recurrent, Transformer, graph-based, spatial-embedding, and language-model approaches.These methods model mobility transitions, geographic context, semantic priors, or textual knowledge.
  • Visual Trajectory Mining: Vision-language trajectory methods provide richer visual context, but their outputs are typically numeric coordinates rather than continuous paths drawn in the visual domain.This distinction motivates the paper’s map-visual generation formulation.
  • Video Generation: Video generation has evolved from VAE and diffusion models toward continuous-time flow-based generators that integrate learned velocity fields.Flow Matching can enable fewer-step sampling and is presented as a substrate for structured refinement.
  • Video Generation: Recent video models motivate using spatiotemporal reasoning to produce topology-consistent continuous paths from map visualizations.The paper draws on demonstrations that video generators can perform spatially related tasks beyond their original training objectives.
  • Reinforcement Learning: RLVR methods extend verifiable-reward optimization to image and video generation by evaluating domain constraints or automatic criteria.The cited approaches use multiple rollouts and group-relative updates to complement supervised training with task-aligned optimization.

3 Preliminaries

The preliminaries define cellular signaling and GPS trajectories, formulate Sig2GPS under map constraints, and describe the flow-based generation foundation. They also explain why stochastic flow sampling supports group-relative reinforcement-learning optimization.

  • Problem Definition: Cellular signaling is represented as a temporal sequence of connections between a mobile phone and stations at spatial coordinates and times.Each record captures a connection to a cellular station located at a specified position during a time interval.
  • Problem Definition: The ground-truth GPS trajectory is an evenly sampled sequence of timestamped coordinates with fixed interval Δt.The formulation assumes ˆt_k+1−ˆt_k=Δt.
  • Problem Definition: Sig2GPS learns a mapping from cellular signaling and map context to a GPS trajectory close to ground truth while remaining spatially plausible under map constraints.The trajectory start time is aligned with signaling, and GPS points are resampled every 15 s.
  • Flow-Based Generation: A conditional flow model integrates a learned velocity field from an initial noise distribution to produce a sample decoded into a trajectory video.The conditioning information can include the signaling visualization.
  • Reinforcement Learning: The deterministic ODE formulation yields identical trajectories under fixed conditions, so an SDE provides stochastic rollouts for group-relative reinforcement-learning updates.The SDE uses a Wiener process and noise schedule, while matching marginal density evolution with a probability-flow ODE under regularity conditions.
  • Reinforcement Learning: The framework uses supervised fine-tuning followed by trajectory-aware RLVR, with group-relative policy optimization applied to sampled rollouts.The training recipe aligns generations with map topology and temporal consistency.

4 Methodology

The framework reframes Sig2GPS as map-visual video generation, then uses trajectory-aware reinforcement learning to align generated paths with distance, direction, continuity, and heterogeneous reward criteria.

  • Training framework: The two-stage recipe first fine-tunes a flow-based video generator on paired signaling-trajectory videos, then applies trajectory-aware reinforcement learning.The reinforcement-learning stage aligns generations with map topology and temporal consistency.
  • Data construction: Training pairs combine cellular signaling with taxi GPS, linked through mobility, time-coverage, and distance-consistency criteria.Taxi GPS serves as the supervision signal, while direct cross-system links are unavailable because of privacy.
  • Map-visual representation: Each pair renders a signaling polyline on an OpenStreetMap tile and progressively draws the ground-truth GPS path as the target video.The model receives the map tile with the signaling trace and generates a video that draws the GPS path over time.
  • Map-visual representation: Map-visual drawing exposes road geometry and mirrors how signaling engineers sketch plausible on-road routes, unlike purely numeric trajectory modeling.Both conditioning and prediction remain in the map-visual domain rather than using an image input with discrete coordinate outputs.
  • Trajectory-aware rewards: Supervised pixel-level objectives can miss wrong turns, reversed direction, and topology violations, motivating rewards that evaluate trajectory-specific errors.The proposed rewards are designed from the evaluation criteria and failure modes observed in the SFT-initialized generator.
  • Trajectory-aware rewards: Distance rewards evaluate geodesic error at five anchor frames and encourage closeness throughout the drawing process.The anchors are the first, last, and three intermediate frames; the resulting distance is mapped to a bounded reward.
  • Trajectory-aware rewards: Direction rewards compare normalized start-to-end displacement vectors, producing R_dir ∈[0, 1], while continuity rewards reject fragmented or multiply-ended paths.The continuity check uses connected components and endpoints under 8-neighborhood connectivity, returning 0 for invalid fragmentation and 1 otherwise.
  • Traj-GDPO optimization: Traj-GDPO normalizes heterogeneous reward components separately before aggregation, addressing instability caused by different ranges, sparsity patterns, and noise levels.The method samples rollout groups, computes group-relative advantages per reward component, aggregates them, and applies batch normalization with clipped policy optimization and KL regularization.

5 Experiments

Experiments show that the video-based Sig2GPS approach consistently outperforms engineered and learning-based baselines, scales with additional temporal context, and transfers across cities and tasks. Ablations and case studies further examine the roles of trajectory-aware rewards, KL regularization, and map-constrained generation.

  • 5.2 Comparison Results on Sig2GPS: Across three trajectory scopes, the proposed model achieves the best overall performance, improving both average reconstruction accuracy and large-error behavior.It records the lowest MAE and RMSE, highest L100, and lowest G1000 across all scopes.
  • 5.2 Comparison Results on Sig2GPS: Processing a single trajectory takes within 30 seconds, compared with more than two minutes for the multi-stage industrial pipeline.The one-step generation paradigm removes explicit post-processing and map-matching stages.
  • 5.3 Test-Time Scaling: Increasing inference frames from 13 to 21 consistently reduces MAE and RMSE, with larger gains for longer trajectory scopes.This test-time scaling requires neither retraining nor architectural changes.
  • 5.4 Comparison Results on Next GPS Prediction and Cross City Transfer: On next-GPS prediction, Think Over Trajectory consistently outperforms both discrete-point reasoning methods and static trajectory-representation methods.The evaluation uses Chengdu and Xi’an benchmarks, including models trained in one city and evaluated in another.
  • 5.4 Comparison Results on Next GPS Prediction and Cross City Transfer: Cross-city transfer is strong: a model trained on Xi’an achieves lower RMSE on Chengdu than all baseline methods.The paper attributes this generalization to structural priors learned from road-network-aligned trajectories.
  • 5.5 Ablation Study: Removing Traj-GDPO or individual rewards degrades performance, while direction reward contributes the majority of gains among the three rewards.Replacing GDPO with GRPO causes a slight performance drop; fixed-reference KL regularization preserves visual fidelity during RL.

6 Conclusion

The paper presents Think Over Trajectory, a map-visual video-generation paradigm for converting coarse cellular signaling into continuous, road-constrained trajectories. Experiments show gains over learning-based and engineered baselines, while qualitative cases support topology-consistent generation.

  • Think Over Trajectory converts coarse cellular signaling traces into continuous, road-constrained trajectories by drawing in the map domain.
  • Traj-GDPO aligns generated trajectories with fine-grained criteria through verifiable rewards after flow-based supervised initialization.
  • Experiments on real-world data demonstrate consistent gains over learning-based baselines and a production-grade engineered pipeline.
  • Qualitative results show topology-consistent routes across diverse scenarios.

A Evaluation Metrics

The evaluation defines point-wise trajectory error using geodesic distance between predicted and ground-truth points, then summarizes these errors with MAE and RMSE.

  • Point-wise error is the geodesic distance between each predicted point and its matched ground-truth point.
  • MAE summarizes the mean absolute point-wise error across matched trajectory points.
  • RMSE summarizes the root mean squared point-wise error across matched trajectory points.

B.1 Sig2GPS

The Sig2GPS evaluation reports trajectory distance statistics for the Sig2GPS dataset. The supplied passages identify the statistics section but do not provide its numerical table entries.

  • The section reports trajectory distance statistics for Sig2GPS evaluation.
  • No numerical distance values or metric comparisons are stated in the supplied passages.
  • The passage identifies a Sig2GPS Dataset statistics subsection.

B.2 Next GPS Prediction

The next-GPS prediction evaluation uses Chengdu and Xi’an subsets from the DiDi Chuxing GAIA Open Dataset. These trajectories are organized as time-ordered GPS observations that capture urban road-network and mobility dynamics.

  • The evaluation adopts Chengdu and Xi’an subsets from the DiDi Chuxing GAIA Open Dataset.
  • The dataset organizes samples at the trip or order level as time-ordered sequences of GPS observations with coordinates and timestamps.
  • The trajectories provide spatiotemporal signals for learning road-topology priors and urban mobility dynamics.

C Baseline Introduction

This section introduces recurrent, convolutional, feed-forward, Transformer, graph-enhanced, agentic, and multimodal baselines for trajectory and mobility modeling. The baselines capture temporal dependencies, spatiotemporal structure, location context, mobility patterns, or flexible trajectory-mining requirements through different mechanisms.

  • Classical sequence baselines: GRU, GPT, MLP, and TCN provide standard sequence-modeling baselines with complementary temporal capabilities.GRU uses update and reset gates; GPT models long-range dependencies with self-attention; MLP uses aggregated or flattened features; and TCN captures local-to-midrange patterns with temporal convolutions.
  • Trajectory Transformers: TrajFormer improves trajectory modeling through continuous point embeddings and squeezed attention for irregular intervals and lower computation.The model explicitly handles irregular spatiotemporal intervals and compresses key/value representations before attention; subpath labeling can provide auxiliary supervision.
  • Signaling-aware modeling: SigFormer reconstructs missing or incorrect base-station locations by using self-attention to model long-range spatiotemporal signaling context.It is designed specifically for mobile signaling data augmentation via location reconstruction.
  • Mobility prediction: DeepMove models sparse, lengthy trajectories with multimodal embeddings, recurrence, and historical attention over mobility periodicity.Its historical attention selectively attends to relevant records to capture multi-level periodic patterns.
  • Contextual and agentic predictors: GETNext, STAN, and AgentMove incorporate global transitions, pairwise spatiotemporal effects, personalized frequency, memory, and external mobility knowledge for location prediction.GETNext uses graph-based POI transitions; STAN models non-adjacent check-in interactions and personalized item frequency; AgentMove combines spatial-temporal memory with world and collective knowledge before reasoning.
  • Multimodal trajectory mining: Traj-MLLM provides training-free, region-agnostic trajectory mining through map-anchored sub-trajectories, interleaved image-text representations, and prompt optimization.The framework uses multi-view context for task adaptation.
Loading 2603.26610v1…