Efficient Dense Crowd Trajectory Prediction Via Dynamic Clustering

arXiv:2603.18166 · cs.AI, cs.CV · Submitted 2026-03-18 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "Efficient Dense Crowd Trajectory Prediction Via Dynamic Clustering".

Tom: Efficient dense crowd trajectory prediction in high-risk environments like transportation hubs requires methods that can handle the challenges of massiveness, noisiness, and inaccuracy inherent in dense crowds.

Jane: First, who's behind it and why it matters.

Title and authors: Tom: Now we're moving into a deeper look at what the authors actually say in the summary of "Efficient Dense Crowd Trajectory Prediction Via Dynamic Clustering." Jane, can you explain in simple terms why they believe this approach is so useful for dealing with the messy crowds we see every day?

Jane: Certainly, Tom; basically, they argue that instead of trying to track every single person individually which gets incredibly hard in dense settings, this method groups them into clusters based on similarities over time <ref:2603.18166#pg0>. This grouping lets the AI focus on summarizing the behavior of a group rather than dealing with thousands of noisy, individual data points at once.

Lu: I think what they are highlighting is that this summarization step is what enables the speed increase; it's like compressing a huge amount of raw data into something meaningful for the prediction model <ref:2603.18166#pg0>. This dynamic grouping process, where members can move in and out of clusters as their movement changes, gives them flexibility that traditional static methods just don't have.

Meng: From an engineering standpoint, that flexibility is crucial because it means the system stays robust even when tracking gets interrupted or some people disappear for a moment <ref:2603.18166#pg0>. The summary emphasizes that this method maintains comparable accuracy to the more intensive state-of-the-art models, which is a big win for deployment on less powerful hardware.

Lalam: And I'm really excited about the aspect of privacy that comes with it; since the final prediction relies on these aggregated cluster representations, we aren't necessarily needing to process every sensitive piece of individual tracking data for the forecasting part <ref:2603.18166#pg0>. This aggregation actually helps build a more trustworthy system culturally because it respects individual privacy while still giving us useful crowd insights.

Tom: That's a huge point about trust, Lalam; moving toward systems that respect privacy while delivering high performance is something the public will really appreciate <ref:2603.18166#pg0>. But Jane, what about the specific metrics they use to prove this works? How do we know these "groups" are actually making good predictions in practice?

Jane: The authors focus heavily on those specific error metrics like CTEO and CTEL, which are designed to measure the naturalness and continuity of the predicted paths <ref:2603.18166#pg0>. They show that trajectories don't have those sudden, jarring jumps that ruin a prediction, which is vital for real-world use.

Lu: And they also have CMDD to look at how far each member is from its center of gravity; that tells us exactly how tightly packed the group behavior is <ref:2603.18166#pg0>. It shows that the directionality component in their features really matters a lot because you can't just throw random people into a cluster if they are all moving in completely different directions <ref:2603.18166#pg0>.

Meng: The results they shared regarding execution time reductions, like that jump from thirty-three point three three percent to seventy-nine point four percent, tells me this isn't just theoretical; it has tangible benefits for low-latency applications <ref:2603.18166#pg0>. That kind of speedup is what makes deploying this technology on edge devices feasible rather than keeping everything in the cloud <ref:2603.18166#pg0>.

Lalam: I think the real impact here, looking at all these factors together, is that we are building a foundation for AI systems that can operate reliably in complex public spaces without overwhelming our privacy concerns or our computational resources <ref:2603.18166#pg0>. This kind of efficiency could lead to a whole new category of monitoring tools <ref:2603.18166#pg0>.

Tom: Exactly, and that leads us right into the big picture, Jane—if we can make this prediction so fast and robust in transportation hubs, what does that actually mean for how society manages massive flows of people?

Jane: It means we can move toward a proactive safety model where AI isn't just reacting to a situation but is predicting potential congestion or unsafe groupings long enough to intervene before an issue escalates <ref:2603.18166#pg0>. This shifts the focus from managing chaos after the fact to managing flow before it happens.

Lu: I see this as opening up entirely new avenues for urban planning and emergency response simulation, because if we can model group dynamics this accurately, we can test different crowd control strategies in a digital environment before implementing them in reality <ref:2603.18166#pg0>.

Meng: Practically speaking, it means we can build systems that are reliable enough for critical infrastructure monitoring where downtime or inaccurate data is simply not an option, which is a huge hurdle right now <ref:2603.18166#pg0>.

Lalam: I think the cultural implication is that as these prediction tools become more reliable and efficient, people will feel safer moving through public areas knowing there's an underlying intelligent system helping to manage the crowd dynamics <ref:2603.18166#pg0>. This is about building trust in automated systems operating in shared physical spaces <ref:2603.18166#pg0>.

The paper's summary: Tom: We’ve discussed the core summary of "Efficient Dense Crowd Trajectory Prediction Via Dynamic Clustering," and now we are looking at the specific technical improvements the authors suggested to make this method even better. Jane, what are some of these refinements they propose?

Jane: The paper suggests refining the centroid calculation by using a delta-based approach every ten frames to smooth out trajectory features, which helps reduce cumulative error from those dynamic membership changes <ref:2603.18166#pg0>. This refinement is crucial for keeping the predicted paths looking natural and continuous.

Lu: Beyond just smoothing, they detail the two main stages of clustering: nested distance-based clustering for direction and distance, followed by a grouping stage that continuously re-evaluates membership based on neighboring clusters <ref:2603.18166#pg0>. That dynamic reassessment is what gives them the flexibility to adapt to changing crowd densities.

Meng: From an engineering standpoint, I find the detail about how they handle new data points—assigning them based on neighboring clusters and putting them into a temporary list if they don't find any nearby groups—to be very important for stability <ref:2603.18166#pg0>. It addresses the dynamic nature of adding and removing members in real-time.

Lalam: And I’m also keen on how they use those specific evaluation metrics like CTEO and CTEL to quantify the quality; they are designed specifically to measure the naturalness of movement, which is a strong indicator of a high-quality prediction <ref:2603.18166#pg0>.

Tom: That’s good—so we have a mechanism for dynamic membership changes and specific metrics to check for path smoothness. Lalam, you mentioned privacy earlier; does this refinement help enhance that aspect of the system?

Jane: It does, because by focusing on aggregating these cluster representations rather than tracking every individual point, the system naturally builds a more robust structure for privacy protection <ref:2603.18166#pg0>. The focus is on group dynamics over raw data points.

Lu: Looking ahead, the authors suggest that their approach can be combined with existing trajectory predictors by using their output centroid in place of the pedestrian's location, which is a practical way to integrate this new idea into current systems <ref:2603.18166#pg0>. That plug-and-play capability is where the real power lies.

Meng: Integrating it into existing predictors means we’re talking about faster deployment because we aren't having to build a whole new system from scratch for every use case, which is a significant practical win <ref:2603.18166#pg0>.

Lalam: I think the implication is that this clustering framework could become a standard way of handling complex crowd data in future AI systems because it balances the need for high accuracy with the constraints of privacy and resource management <ref:2603.18166#pg0>.

The paper's improvements: Tom: We've spent our time walking through the "Efficient Dense Crowd Trajectory Prediction Via Dynamic Clustering" paper, and I think we’re ready to wrap up our discussion on this interesting research. Jane, can you give us the final summary of what this study actually achieved?

Jane: Well, Tom, the main conclusion is that by grouping individuals based on their changing attributes over time using a dynamic process, they manage to maintain high prediction accuracy while significantly lowering the computational cost and memory requirements for these dense crowd tasks <ref:2603.18166#pg0>.

Lu: It's really neat how they managed to balance that speed increase with keeping the prediction quality up, especially when you factor in those specific metrics like CTEO and CTEL that they used to check for path smoothness <ref:2603.18166#pg0>.

Meng: I agree, Lu; the practical result is a system that's much more deployable because it doesn't require massive computational power for every single frame of dense crowd tracking <ref:2603.18166#pg0>. That reduction in resource usage is what makes this approach viable for real-time applications.

Lalam: I think the bigger vision here is that we're developing a framework where complex, massive data can be summarized into manageable group dynamics, which could fundamentally improve how we design and interact with large-scale public safety and monitoring systems <ref:2603.18166#pg0>.

Tom: That's a powerful way to put it, Lalam; moving from raw data overload to intelligent abstraction for better societal outcomes. Jane, what's the final thought on how this work might shape the future of trajectory prediction?

Jane: I think this research sets a solid foundation because it shows that simplifying complex input data through dynamic grouping is a viable path toward more efficient and robust AI models <ref:2603.18166#pg0>. It paves the way for handling environments where tracking individual points is simply impractical.

Lu: Looking ahead, I see this as a starting point for exploring even richer cluster representations; we could potentially expand on how those group attributes are learned to make them even more predictive over longer time horizons <ref:2603.18166#pg0>.

Meng: For me, the future impact is about making these kinds of efficient prediction tools accessible across a wider range of industries, not just transportation hubs, because if we can get this efficient on edge devices, it opens up a ton of other areas <ref:2603.18166#pg0>.

Lalam: Ultimately, the advancement in Efficient Dense Crowd Trajectory Prediction Via Dynamic Clustering allows us to build AI systems that are more effective at managing our shared physical world, which is a huge step for how we experience public safety and movement <ref:2603.18166#pg0>.

Conclusion: Tom: So we've seen how they use dynamic clustering to group pedestrians based on their movement patterns over time to make prediction faster and more accurate, which is what they call "Efficient Dense Crowd Trajectory Prediction Via Dynamic Clustering." Now, Jane, can you give us the final summary of what this study actually achieved?

Meng: I agree, Lu; the practical result is a system that's much more deployable because it doesn't require massive computational power for every single frame of dense crowd tracking. That reduction in resource usage is what makes this approach viable for real-time applications.

Jane: I think this research sets a solid foundation because it shows that simplifying complex input data through dynamic grouping is a viable path toward more efficient and robust AI models. It paves the way for handling environments where tracking individual points is simply impractical.

Tom: We've covered so much today regarding the "Efficient Dense Crowd Trajectory Prediction Via Dynamic Clustering" paper, and I think we're ready to wrap up our discussion on this interesting research. Jane, can you give us the final summary of what this study actually achieved?

Meng: I agree, Lu; the practical result is a system that's much more deployable because it doesn't require massive computational power for every single frame of dense crowd tracking. That reduction in resource usage is what makes this approach viable for real-time applications.

Jane: I think this research sets a solid foundation because it shows that simplifying complex input data through dynamic grouping is a viable path toward more efficient and robust AI models. It paves the way for handling environments where tracking individual points is simply impractical.

University of Glasgow

cs.AI, cs.CV

Submitted: 2026-03-18

Updated: 2026-10-08

Code: https://github.com/bimamurti/CrowdCluster

Importance score: 80/100

The gist: Efficient dense crowd trajectory prediction in high-risk environments like transportation hubs requires methods that can handle the challenges of massiveness, noisiness, and inaccuracy inherent in

Key concepts

Nested Distance-Based Clustering: This is the initial step where the system groups pedestrians based on their location and direction. It uses Local Outlier Factor (LOF) to identify outliers, ensuring that similar movement patterns are grouped together effectively. This helps in reducing noise from tracking data.
Centroid Calculation with Delta-Based Approach: Instead of calculating the centroid directly every frame, this method updates it by averaging the differences between consecutive locations over 10 frames. This delta approach smooths out cumulative errors caused by dynamic membership changes, resulting in more stable and accurate cluster trajectory representations.
Cluster Trajectory Errors Occurrence (CTEO): This metric measures how natural the predicted paths are for each group. It counts the percentage of noticeable path deviations within a cluster's trajectory. A low CTEO score indicates that the resulting movement patterns are smooth and continuous, which is crucial for realistic crowd prediction.

Terminology

Clustering — 85× in this paper · explained in 2 episodes
Accuracy — 32× in this paper · explained in 1 episode
Computer Vision and Pattern Recognition — 5× in this paper
Edge Devices — 4× in this paper · explained in 2 episodes
Abstraction — 3× in this paper · explained in 1 episode
Transformer — 3× in this paper · explained in 2 episodes
Cluster Member Distance Deviations — 2× in this paper
Cluster Trajectory Errors Length — 2× in this paper
Cluster Trajectory Errors Occurrence — 2× in this paper
Generative Model — 2× in this paper · explained in 2 episodes
Local Outlier Factor — 2× in this paper
Optical Flow — 2× in this paper · explained in 2 episodes
Aggregation — 1× in this paper · explained in 1 episode
Benchmark — 1× in this paper · explained in 2 episodes
Benchmarking — 1× in this paper · explained in 3 episodes
Core — 1× in this paper · explained in 2 episodes
Distillation — 1× in this paper · explained in 2 episodes
Long Short-Term Memory — 1× in this paper
Robustness — 1× in this paper · explained in 5 episodes
Unified Framework — 1× in this paper · explained in 5 episodes

Summary

Efficient dense crowd trajectory prediction in high-risk environments like transportation hubs requires methods that can handle the challenges of massiveness, noisiness, and inaccuracy inherent in dense crowds. This work proposes a novel cluster-based approach that groups individuals based on similar attributes over time to enable faster execution through accurate group summarization, leading to faster processing and lower memory usage while maintaining comparable accuracy compared to state-of-the-art methods.

The gist

Our proposed method groups individuals into clusters based on similar attributes over time, enabling faster execution through accurate group summarization.

How it works

The proposed dynamic clustering method inputs a set of time-varying pedestrian locations, typically from an existing tracker. The process consists of two main stages: nested distance-based clustering for direction and distance, followed by a grouping stage that continuously re-evaluates cluster membership. This approach reduces training and inference time and memory usage for the prediction model, while being robust to noisy tracking input.

The method is based on two main input features: location (x, y), direction angle (θ), direction vector (vx, vy), and cluster id (Cid). The direction vector is calculated by subtracting the previous and current location vectors, and the direction angle is the arctangent of its components.
The clustering process begins 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, they are assigned to another cluster nearby or stored on an unassigned list.
The algorithm employs a dynamic process for adding and removing members: New data is also assigned based on neighbouring clusters and put into the temporary list if they do not find any clusters nearby. This handles the adding and removing process of dynamic clustering.

Centroid Calculation

The crucial step for forming cluster trajectories is determining the centroid feature values. The centroid vector feature is determined as:

Cit = Xit; Yit; θit, where C represents the centroids, i represents the cluster number, and t represents the frame number.

To ensure smooth trajectories and reduce cumulative error from dynamic membership changes, a delta-based approach is used for centroid calculation. Every 10 frames, the centroid's location is determined by averaging the differences between current and previous locations:

∆Cpit = Pn j=0(Cpjt − Cpjt−1) n, leading to Cpit = Cpit−1 + ∆Cpit. The direction vector is then calculated using this new location to reduce cumulative error.

Evaluation Metrics

The paper defines several evaluation metrics to quantify the performance of the clusters and their resulting trajectories:

(CTEO)

Cluster Trajectory Errors Occurrence (CTEO): Pedestrian trajectories should appear natural, meaning that no sudden displacements appear in their paths, which could disrupt the trajectory continuity. This is calculated by counting the percentage value of the noticeable path deviation occurrence of every cluster’s trajectories.

(CTEL)

Cluster Trajectory Errors Length (CTEL): The Length of error shows how much the error affects the entire cluster’s trajectory. This counts every cluster trajectory error length and divides it by the number of clusters.

(CMDD)

Cluster Member Distance Deviations (CMDD): This metric calculates the average value of the distance from every cluster member to its centroid, and calculates the average value for all clusters and frames. It shows how far the member is from its centroid, with a note that The direction took a more crucial part since pedestrians can not be in a cluster if they move in a different direction.

Performance Results

The experiments demonstrate that the approach maintains prediction accuracy while significantly reducing computational cost. For instance, when evaluated on Trajectron++, SocialVAE, and MART, the proposed clustering approach significantly reduced execution time (from 33.33% to 79.4%) and reduced its maximum memory usage up to a certain level compared to using raw pedestrian data or traditional methods. Furthermore, in long-term prediction scenarios, the clustering approach surpasses the accuracy of this random selection method. The results indicate that the cluster trajectory is smooth without any anomalous displacement of the trajectory from one location to another.

Conclusion

The novel dynamic clustering approach successfully reduces execution time and memory usage while maintaining accuracy with minimal performance degradation on short-term prediction and even performance improvement on long-term prediction compared to the tracking input and random selection. This research lays the groundwork for future investigations into real-time prediction by simplifying dense crowd trajectory data.

Acknowledgements

This research was funded by the Center of Higher Education Funding and Assessment (PPAPT), the Indonesian Ministry of Higher Education and Research, the Indonesian Education Scholarship (BPI), and the Indonesian Endowment Fund for Education (LPDP).

References

Improvements for AI systems

Here are the specific improvements that can be made to existing AI trajectory prediction systems based on the proposed Dynamic Clustering approach, and what those improved systems can achieve:

The core improvement lies in replacing individual pedestrian tracking inputs with dynamically evolving cluster centroids, which significantly reduces computational load while maintaining predictive accuracy in dense crowds.

Enhancement of Computational Efficiency (Inference Speed and Memory Footprint):
Reduction of Tracking Noise Sensitivity:
Improved Robustness to Missing/Switching Identities:
Lower Latency in Real-Time Applications:

Specific Capabilities of the Improved AI System:

Predict trajectories in high-density environments (e.g., large stadiums, busy transit hubs) with significantly faster inference times (up to a 79% reduction in execution time compared to raw data processing) and lower memory consumption, making real-time deployment on edge devices feasible where current state-of-the-art models are too resource-intensive.
Maintain high prediction accuracy (comparable ADE/FDE scores) even when the underlying pedestrian tracking system suffers from common issues like ID switches, tracking loss, or noisy head detection in dense scenes, because the model relies on robust group dynamics rather than individual noisy points.
Enable more reliable long-term trajectory forecasting (up to 50 steps ahead) by effectively summarizing group behavior over time through the centroid calculation method, outperforming naive methods like random trajectory sampling and random selection.
Provide a privacy-preserving input mechanism for trajectory prediction, as the system uses aggregated cluster data instead of sensitive individual tracking data for the final forecasting step.
Improve path smoothness and continuity in predicted trajectories (as quantified by low CTEO and CTEL metrics), ensuring that the predicted movements are physically plausible and do not exhibit abrupt, unnatural displacements caused by transient tracking errors.

Sources

CrowdHuman: A Benchmark for Detecting Human in a Crowd