Efficient Dense Crowd Trajectory Prediction Via Dynamic Clustering

Antonius Bima Murti Wijaya; Paul Henderson; Marwa Mahmoud

cs.AIMar 18, 2026

Efficient Dense Crowd Trajectory Prediction Via Dynamic Clustering

Authors: Antonius Bima Murti Wijaya, Paul Henderson, Marwa Mahmoud

Organizations: University of Glasgow, Glasgow G12 8QQ, United Kingdom

Abstract

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

Figures & tables

Figure 1: Our proposed dynamic clustering for dense crowds. The process starts with initialisation with a nested agglomerative cluster, evaluating the direction every 10 frames with LOF (Local Outlier Factor), and calculating the centroids. If LOF identifies outliers, then each outlier will be assigned to another nearby cluster or stored on an unassigned list. The process will execute the nested cluster again when a member of the unassigned list reaches a certain value. The Centroid Trajectory Calculation section shows how the cluster calculates its centroid trajectory accelerations based on the average deviation value from its membership value. The Cluster Assignment section illustrates the dynamic processing of clustering by the algorithm.

Scene	IDF1 $\uparrow$	Rcll $\uparrow$	MOTA $\uparrow$	FN $\downarrow$	IDs $\downarrow$
02	61.2 %	74.2 %	66.8%	168313	1890
03	43.9%	56.2%	51.3%	103562	2144
04	49.14%	60.4%	53.67%	61286	919

Table 1: Tracking evaluation for MOT HT21 Dataset. The table shows the results of tracking error. The least noise was obtained on the HT21-02 and the most noise was obtained on the HT21-04 scenario.

Data	CMDD $\downarrow$	CTEO $\downarrow$	CTEL $\downarrow$	ncluster	npeds
HT-02	7.93	0.0	0.0	308/460	1264
GT-02	6.77	0.26	0.01	267/378	602
HT-03	7.04	0.23	0.01	443/882	2577
GT-03	5.60	0.98	0.02	353/482	811
HT-04	7.99	0.31	0.01	281/653	1387
GT-04	8.084	0	0.0	270/367	580

Table 2: Cluster evaluation based on automated tracking outputs (HT) and ground truth trajectories (GT). All HT scenarios exhibit consistently low CMDD values, indicating minimal average distance across all frames. The ncluster column shows the number of clusters created between one member and more than 2 members, and npeds shows the total number of pedestrians.

Scene	Source	Exec $\downarrow$	ADE $\downarrow$	FDE $\downarrow$	Train Data	nNode
02	GT	2703.3	3.06	6.53	128,217	448
02	Cluster	672.4	9.74	17.72	62,396	268
02	Tracking	1496.5	9.56	17.40	102,852	805
03	GT	1036.2	25.90	47.63	161,198	616
03	Cluster	530.0	26.50	48.69	60,539	528
03	Tracking	2572.0	26.41	48.66	110,729	1599

Table 3: Trajectory prediction performance using Trajectron++ on HeadHunter.

Scene	Source	Exec $\downarrow$	A12 $\downarrow$	A50 $\downarrow$	F12 $\downarrow$	F50 $\downarrow$	Mem $\downarrow$
02	GT	23.3	1.83	9.90	3.15	17.23	70.26
02	Cluster	13.3	2.02	9.51	3.35	16.89	23.64
02	Tracking	20.0	2.01	12.18	3.43	20.86	45.72
02	Random	13.3	-	13.47	-	23.44	-
03	GT	30.0	3.20	15.97	4.802	34.08	120.17
03	Cluster	13.3	4.10	18.66	6.79	37.11	17.85

Table 4: Trajectory prediction performance with SocialVAE. This table shows a significant execution time reduction and shows how the cluster outperforms the accuracy of the tracking results on 50 future steps and traditional random selection methods.

Scene	Source	Exec $\downarrow$	A12 $\downarrow$	A50 $\downarrow$	F12 $\downarrow$	F50 $\downarrow$	Mem $\downarrow$
02	GT	748.4	1.12	5.56	1.95	8.61	33.92
02	Cluster	236.3	1.55	6.22	3.03	9.78	15.79
02	Tracking	314.2	1.51	5.99	2.36	9.32	25.30
02	Random	136.5	-	6.37	-	10.16	14.81
03	GT	1485.2	1.89	6.59	2.72	9.64	54.66
03	Cluster	167.3	2.82	9.77	4.22	16.18	20.69

Table 5: Trajectory prediction performance with MART. This table shows a significant reduction in execution time and shows how the cluster outperforms the execution time and memory usage of the tracking results and traditional random selection methods.

Scene	Source	Exec $\downarrow$	A12 $\downarrow$	A50 $\downarrow$	F12 $\downarrow$	F50 $\downarrow$	Mem $\downarrow$
02	GT	485.2	1.21	6	2.32	10.24	6.40
02	Cluster	126.2	1.65	6.632	3.50	11.288	2.98
02	Tracking	183.4	1.62	6.42	2.77	10.952	4.78
02	Random	84.0	-	10.648	-	18.592	2.80
03	GT	961.5	2.75	9.6	4.78	16.936	10.32
03	Cluster	161.3	3.10	11.584	4.74	21.512	3.91

Table 6: Trajectory prediction performance with DSTIGCN. The table shows a significant reduction in execution time and Memory from the tracking result, and a slight reduction or increase in accuracy for both prediction horizons.

Pedestrian Trajectory Prediction

Pedestrian trajectory prediction forecasts where people will move from their observed motion and surrounding context.

See all 36 papers

CrowdTraj: A Benchmark for Dense Crowd Trajectory Prediction in Realistic Crowded Environments
Sep 7, 2026Antonius Bima Murti Wijaya, Paul Henderson, Marwa MahmoudPedestrian Trajectory Prediction Multi-Object Tracking
School of Computing Science University of Glasgow University Avenue G12 8QQ, Glasgow, United Kingdom
Social-Mamba: Socially-Aware Trajectory Forecasting with State-Space Models
May 14, 2026Po-Chien Luan, Wuyang Li, Yang Gao +1Pedestrian Behavior Modeling Pedestrian Trajectory Prediction
EPFL, Switzerland
STDDN: A Physics-Guided Deep Learning Framework for Crowd Simulation
Apr 3, 2026Zijin Liu, Xu Geng, Wenshuai Xu +3Pedestrian Trajectory Prediction Physics-Based Simulation
School of Computer Science and Engineering, Beihang University, China · Beijing Digital Native Digital City Research Center, Beijing, China · School of Software, Beihang University, China

Sep 7, 2026J/K move · Enter open · S save

CrowdTraj: A Benchmark for Dense Crowd Trajectory Prediction in Realistic Crowded Environments

Antonius Bima Murti Wijaya, Paul Henderson, Marwa Mahmoud

School of Computing Science University of Glasgow University Avenue G12 8QQ, Glasgow, United Kingdom

Pedestrian Trajectory Prediction Multi-Object Tracking

Open paper AskPDF arXiv

In real-world applications, pedestrian trajectory prediction models rely on inputs from detection and tracking systems. Prior trajectory prediction benchmarks either contain relatively sparse pedestrian interactions, assume perfect tracking inputs, or rely on overhead viewpoints that minimize occlusion and perspective distortion, limiting evaluation in realistic dense-crowd scenarios. We present CrowdTraj, a benchmark for pedestrian trajectory prediction in natural dense crowd scenes. Unlike previous datasets, CrowdTraj supports end-to-end evaluation from detection through tracking to trajectory prediction under severe occlusion in CCTV views. It also captures diverse, natural pedestrian behaviours, including abrupt directional changes rarely observed in existing benchmarks. CrowdTraj includes five diverse scenes, with an average of 1,146 unique pedestrians per scene, maximum frame-level densities ranging from 114 to 372 pedestrians, and over 3.2 million annotated head bounding boxes. CrowdTraj provides pixel and real-world coordinates via per-scene homography matrices for physically meaningful analysis. Our experimental results show that tracking accuracy (IDF1) drops to 0.68 to 0.70 in the densest scenes, compared with approximately 0.90 in less crowded scenes. Trajectory prediction training also becomes substantially more computationally expensive in dense scenes, with training times increasing by up to 8 times. These findings show that CrowdTraj exposes limitations in current trajectory prediction pipelines that remain hidden on existing sparse-crowd benchmarks, particularly in robustness to tracking noise and computational scalability.

Abstract

Pedestrian Trajectory Prediction

Explore similar work

CrowdTraj: A Benchmark for Dense Crowd Trajectory Prediction in Realistic Crowded Environments

Social-Mamba: Socially-Aware Trajectory Forecasting with State-Space Models

STDDN: A Physics-Guided Deep Learning Framework for Crowd Simulation