PaperlayerView exact revision on arXiv

Source-linked AI summary

Efficient Dense Crowd Trajectory Prediction Via Dynamic Clustering

Antonius Bima Murti Wijaya, Paul Henderson, Marwa Mahmoud

arXiv:2603.18166v1cs.AIcs.CV

TL;DR

Dense-crowd prediction is challenging because prior work often relies on manually annotated tracking data, while tracking outputs can be noisy and costly to process. This paper replaces individual pedestrian inputs with dynamic cluster centroids, reporting faster execution and lower memory use while maintaining accuracy. Its clustering method may produce single-member clusters when pedestrians share a direction but are far apart.

  • Problem

    Prior work focused mainly on lower-density scenarios and manually annotated tracking data; in dense crowds, reduced tracking accuracy affects prediction accuracy while computational cost increases.

    Source: §1 · p. 1

  • Method

    Dynamic clustering groups pedestrians and supplies cluster centroids in place of individual pedestrian inputs to existing trajectory predictors.

    Source: §1 · p. 2

  • Results

    The method reduces execution time and memory usage while maintaining accuracy, with minimal degradation for short-term prediction and improvement for long-term prediction versus tracking input and random selection.

    Source: §1 · p. 2, §5 · p. 13

  • Takeaways & Limitations

    Aggregated cluster data reduces privacy leakage from tracking inputs.

    Source: §1 · p. 2

    Clusters could contain a single member when a pedestrian moves in the same direction as others but is far away.

    Source: §4.4 · p. 12

Abstract

from arXiv · show

Crowd trajectory prediction plays a crucial role in public safety and management, where it can help prevent disasters such as stampedes. Recent works address the problem by predicting individual trajectories and considering surrounding objects based on manually annotated data. However, these approaches tend to overlook dense crowd scenarios, where the challenges of automation become more pronounced due to the massiveness, noisiness, and inaccuracy of the tracking outputs, resulting in high computational costs. To address these challenges, we propose and extensively evaluate a novel cluster-based approach that groups individuals based on similar attributes over time, enabling faster execution through accurate group summarisation. Our plug-and-play method can be combined with existing trajectory predictors by using our output centroid in place of their pedestrian input. We evaluate our proposed method on several challenging dense crowd scenes. We demonstrated that our approach leads to faster processing and lower memory usage when compared with state-of-the-art methods, while maintaining the accuracy

1 Introduction

The paper proposes dynamic clustering that replaces individual pedestrian inputs with cluster centroids for trajectory prediction. It reports faster execution and lower memory usage while maintaining comparable accuracy.

  • Prior work focused mainly on lower-density scenarios and manually annotated tracking data, while dense crowds make tracking less accurate and prediction more computationally costly.Source: §1 · pp. 1–2
  • The method groups pedestrians into clusters and uses their centroids as predictor inputs, reducing sensitivity to disappearance caused by re-identification failures.Aggregated cluster data also reduces privacy leakage from tracking inputs.Source: §1 · p. 2
  • The method is combined with several trajectory predictors and evaluated for cluster membership accuracy and prediction performance on challenging dense-crowd datasets.Source: §1 · p. 2
  • The paper identifies dynamic pedestrian clustering, novel metrics for group-dynamics summarisation, and evaluation on dense-crowd datasets as its three technical contributions.Source: §1 · p. 2
  • The approach achieves faster execution and lower memory usage while maintaining comparable accuracy to state-of-the-art and traditional approaches.Source: §1 · p. 2

2 Related Work

Related work covers pedestrian tracking, trajectory prediction, and dynamic clustering. It notes persistent tracking noise, computational costs in dense crowds, and a need to handle members leaving clusters.

2.1 Pedestrian Tracking

  • Detection-based tracking methods, including SORT [3] and OC-Sort, still face identity switches and tracking loss, especially on low-resolution data.Source: §2.1 · p. 3

2.2 Pedestrian Trajectory Prediction

  • Social and group-aware trajectory methods improve prediction, but per-pedestrian neighbor-grid processing may make dense-crowd training excessive; Grouptron’s cluster-derived group data improved ADE and FDE performance while adding computational load.SocialVAE stores all neighborhood data at initialization and slows prediction because it clusters outputs; Groupnet [27] provides pedestrian grouping, and MART extends Groupnet with a transformer-based model.Source: §2.2 · pp. 3–4

2.3 Dynamic Clustering for Pedestrians

  • Dyclee uses distance- and density-based clustering, but does not evaluate whether members leave a cluster as pedestrian trajectories or velocities change.The paper identifies a need for an additional step to address this dynamic behavior.Source: §2.3 · p. 4

3 Method

The method groups pedestrians dynamically and supplies cluster-centroid trajectories to existing predictors. It combines direction- and location-based clustering with centroid updates designed to preserve smooth trajectories as membership changes.

3 Method

  • Cluster centroids replace individual pedestrian positions as predictor inputs, reducing training and inference time and memory while tolerating noisy tracking input.The method takes time-varying pedestrian locations, typically from an existing tracker, and produces a centroid location for each cluster in every frame.Source: §3 · p. 4

3.1 Clustering Strategy

  • Nested agglomerative clustering first groups pedestrians by direction, then by location, constraining clusters to similar directions and locations within a distance threshold.The approach uses direction angle and location as its two main input features.Source: §3.1 · p. 5
  • Every 10 frames, LOF reevaluates cluster membership; anomalies join a nearby cluster or wait unassigned until the nested clustering process runs again.The initial clusters are insufficient for longer scenarios as pedestrian locations, directions, and numbers change over time.Source: Figure 1 · p. 5, §3.1 · p. 6
  • The trajectories shown for HT21 scenes 2–4 retain overall patterns despite fewer agents, which reduces memory usage and execution time.The trajectories cover frames 70–400, and their occurrence can vary by frame.Source: Figure 2 · p. 7, §3.1 · p. 7

3.2 Centroid Calculation

  • Every 10 frames, centroid locations are updated by averaging members’ displacements rather than averaging their current positions, avoiding abrupt shifts from changing membership and tracking noise.The direction vector and angle are recalculated from the newest centroid location to reduce cumulative error; the resulting deviation mean is smoother than without the method.Source: §3.2 · p. 8

4 Experiments

The experiments assess clustering quality on ground-truth and tracker-generated trajectories, then test centroid inputs with trajectory predictors. They report clustering quality, tracking-data conditions, evaluation measures, and prediction performance across several settings.

4 Experiments

  • The experiments first evaluate clustering as a representation of pedestrian behaviour, then test centroid-based inputs with Trajectron++, SocialVAE, and MART.The clustering evaluation compares ground-truth trajectories with automated HeadHunter-T tracking outputs.Source: §4 · p. 8

4.1 Evaluation Metrics

  • Evaluation combines CTEO path-deviation occurrence, CTEL deviation length, CMDD member-to-centroid distance for clusters with more than two members, and predictor ADE/FDE.The metrics are intended to assess cluster representation, tracking robustness, trajectory smoothness, and prediction accuracy.Source: §4.1 · pp. 8–10

4.2 Implementation Details

  • The tested clustering used 120 px and 50° thresholds; ADE/FDE used 8-step history and 12-step prediction, with 25/50-step settings for SocialVAE, MART, and random selection.Standard deviations were calculated across 10 random-seed repetitions for Trajectron++ and SocialVAE, but not MART.Source: §4.2 · p. 10

4.3 Datasets and Track Extraction

  • HT21 tracking was least noisy on HT21-02 and most noisy on HT21-04, and the reported tracking metrics remained far from perfect.The text says HT21-02 had the highest correct-detection, recall, and MOTA ratios, while false negatives and ID switches were mostly in HT21-04.Source: Table 1 · p. 11, §4.3 · p. 11

4.4 Clustering Quality Results

  • GT and HeadHunter clusters showed similar membership representation; tracking loss and re-identification yielded more pedestrians and clusters in HT, while cluster trajectories had very low CTEL and CTEO.Table 2 reports consistently low CMDD values across HT scenarios.Source: §4.4 · p. 11, Table 2 · p. 11

4.5 Trajectory Prediction Results

  • Across trajectory-prediction results, clustering reduced execution time and memory usage versus raw pedestrian data, with varied slight accuracy changes, and its accuracy surpassed random selection in all cases.Source: §4.5 · p. 13

5 Conclusion

The conclusion reports that dynamic clustering reduces execution time and memory usage while maintaining prediction accuracy, with minimal short-term degradation and improved long-term performance. It presents the approach as groundwork for future real-time prediction by simplifying dense-crowd trajectory data.

5 Conclusion

  • The approach significantly reduces execution time and memory usage, with minimal short-term prediction degradation and improved long-term performance versus tracking input and random selection.Source: §5 · p. 13
  • Dynamic membership distance supports accurate pedestrian representation while maintaining trajectory smoothness and continuity.Source: §5 · p. 13
  • The approach uses nested clustering for pedestrian direction and distance, followed by member evaluations, as groundwork for future real-time prediction.The conclusion frames this as simplifying dense-crowd trajectory data.Source: §5 · p. 14
Loading 2603.18166v1…