TrajFlow: A Generative Framework for Occupancy Density Estimation Using Normalizing Flows
Mitch Kosieradzki Department of Computer Science University of Minnesota [email protected] Seongjin Choi, Ph.D. Department of Civil, Environmental, and Geo- Engineering University of Minnesota [email protected] · 2025
Read TrajFlow: A Generative Framework for Occupancy Density Estimation Using Normalizing Flows by Mitch Kosieradzki Department of Computer Science University of Minnesota [email protected] Seongjin Choi, Ph.D. Department of Civil, Environmental, and Geo- Engineering University of Minnesota [email protected] (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-14266.txt.
Introduction
Predicting where traffic agents will move next is the aim of TrajFlow, a 2025 system by Mitch Kosieradzki and Seongjin Choi. Cars, bikes, and walkers move in uncertain ways, so planners need a map of likely future positions rather than one fixed path. TrajFlow estimates that spread of likely positions directly.
An encoder first turns an observed path into a compact summary of its meaning. A normalizing flow then turns that summary into a density over future positions at a chosen time. Working with the spread itself, rather than full hidden paths, raised accuracy on hard test sets. The spread can also be sampled freely and converted into paths or grid maps for planning.
Catalog record
| CREATOR | Mitch Kosieradzki Department of Computer Science University of Minnesota [email protected] Seongjin Choi, Ph.D. Department of Civil, Environmental, and Geo- Engineering University of Minnesota [email protected] |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 48 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-14266
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
TrajFlow: A Generative Framework for Occupancy Density Estimation Using Normalizing Flows Mitch Kosieradzki Department of Computer Science University of Minnesota [email protected] Seongjin Choi, Ph.D. Department of Civil, Environmental, and Geo- Engineering University of Minnesota [email protected] Abstract \AuthorHeaders Kosieradzki and Choi \TotalWords 6013 1 Abstract For intelligent transportation systems and autonomous vehicles to operate safely and efficiently, they must reliably predict the future motion and trajectory of surrounding agents within complex traffic environments. At the same time, the motion of these agents is inherently uncertain, making accurate prediction difficult. In this paper, we propose TrajFlow , a generative framework for estimating the occupancy density of dynamic agents. Our framework utilizes a causal encoder to extract semantically meaningful embeddings of the observed trajectory, as well as a normalizing flow to decode these embeddings and determine the most likely future location of an agent at some time point in the future. Our formulation differs from existing approaches because we model the marginal distribution of spatial locations instead of the joint distribution of unobserved trajectories. The advantages of a marginal formulation are numerous. First, we demonstrate that the marginal formulation produces higher accuracy on challenging trajectory forecasting benchmarks. Second, the marginal formulation allows for fully continuous sampling of future locations. Finally, marginal densities are better suited for downstream tasks as they allow for the computation of per-agent motion trajectories and occupancy grids, the two most commonly used representations for motion forecasting. We present a novel architecture based entirely on neural differential equations as an implementation of this framework and provide ablations to demonstrate the advantages of a continuous implementation over a more traditional discrete neural network based approach. The code is available at https://github.com/UMN-Choi-Lab/TrajFlow . Keywords : Motion Forecasting, Intelligent Transportation Systems, Autonomous Vehicles, Neural Differential Equations, Normalizing Flows 2 Introduction For intelligent transportation systems (ITS) and autonomous vehicles (AVs) to operate safely and efficiently, they must reliably predict the future motion and trajectory of surrounding agents within complex traffic environments. Accurate trajectory prediction is a fundamental ITS problem, which plays a key role in enabling a wide range of ITS subsystems aimed at improving mobility, safety, and efficiency in transportation networks [ 1 ] . For instance, accurate trajectory forecasting enhances Advanced Driver Assistance Systems (ADAS) by enabling functionalities such as lane-keeping, adaptive cruise control, and proactive collision avoidance through the anticipation of nearby vehicles’ behavior. Within Advanced Traffic Management Systems (ATMS), such predictions enhance vehicle-level control strategies at intersections and during congestion, where precise interactions between agents must be coordinated. Moreover, as ITS expands to include connected infrastructure, predictive models at the microscopic level will also improve real-time routing, intersection control, and pedestrian safety measures. However, predicting future agent trajectories remains challenging due to inherent uncertainties and the stochastic nature of human and vehicular movements in transportation networks. Early attempts at trajectory prediction, such as Social LSTM [ 2 ] , SR-LSTM [ 3 ] , and Social Attention [ 4 ] , focused on deterministically forecasting a single future trajectory for each participant. These deterministic regressors struggled to capture the inherent uncertainty in future motion prediction. To address this limitation, research has shifted towards probabilistic forecasting . For example, Pajouheshgar et al. [ 5 ] utilized spatio-temporal convolutional neural networks (CNNs) [ 6 ] to predict multiple future trajectories probabilistically. Trajectron++ [ 7 ] combined spatio-temporal CNNs with recurrent neural networks (RNNs) for multi-modal forecasting, while FloMo [ 8 ] employed RNNs and normalizing flows [ 9 ] for trajectory density estimation. Normalizing flows have emerged as a promising probabilistic framework due to their ability to compute exact likelihoods through sequence of invertible transformations. Unlike other probabilistic models that rely on predefined distributions (e.g., Gaussian Mixtures [ 10 ] ) or other generative models like Variational Auto-Encoders (VAEs) [ 11 ] and Generative Adversarial Networks (GANs) [ 12 ] that only rely on approximate inference or implicit density modeling, normalizing flows allow for direct likelihood estimation with efficient sampling. This property makes them particularly well-suited for safety-critical applications. In such a context, the ability to explicitly model probability distribution enables robust decision-making under quantifiable uncertainty, enhances risk assessment, and supports scenario planning for rare or dangerous events. Figure 1: Trajectory forecasts of the TrajFlow framework when the model is trained to learn the joint distribution (left) and marginal distribution (right). Yet, while probabilistic forecasting, including normalizing flows, has made significant progress, important challenges remain in how uncertainty is modeled and leveraged in practice. Many existing approaches primarily focus on modeling the joint density of these locations (as shown in Figure 1 (a)) and often operate under a fixed and discrete forecast horizon. Although the joint formulation facilitates the sampling of smooth future motion trajectories and maintains diversity in the samples, obtaining marginal spatial distributions from the joint density is computationally intractable. In contrast, marginalized spatial densities (as shown in Figure 1 (b)) allow for an infinite and continuous forecast horizon, offering significant benefits for downstream tasks by supporting both trajectory reasoning and occupancy-based representations [ 13 ] . Integrating these representations allows systems to maintain comprehensive situational awareness, which is essential for solving complex planning and control challenges. For example, in autonomous vehicles, these representations help intelligent agents plan safe, collision-free motion by predicting the future motion and occupancy of surrounding vehicles. Similarly, in transportation networks, traffic accidents are typically concentrated in areas known as black-spots [ 14 , 15 ] . Previous work has explored the use of the Poisson-Tweedie model [ 16 ] , screening [ 17 ] , clustering [ 17 ] , and crash prediction [ 17 ] as black-spot identification methods. However, no standardized method for black-spot identification has yet been proposed that applies to all road types [ 15 ] . Marginal spatial density estimations facilitate the computation of collision probabilities, enabling intelligent transportation systems to detect black-spots for a wide variety of road types automatically. Additionally, many existing methods predominantly use discrete neural networks in the construction of their forecasting models. However, trajectory data is inherently continuous, and discrete modeling approaches often fail to fully capture all of the information present in such data. One way to address this limitation is by leveraging neural differential equations [ 18 , 19 ] in the model’s construction. Neural differential equations are a continuous modeling paradigm that aligns with the continuous nature of trajectory data. In this work, we introduce TrajFlow , a novel framework for the estimation of probabilistic occupancy densities using normalizing flows (see Figure 2 for examples). Our framework directly models the marginal occupancy density for each future agent location, rather than the joint distribution over future motion trajectories. TrajFlow employs a normalizing flow conditioned on an embedding produced by a causal encoder to model these densities. We explore both discrete and continuous implementations of the TrajFlow framework and provide ablation studies to demonstrate the performance advantages of the continuous approach. By combining neural differential equations with a marginal density formulation, TrajFlow enables a fully continuous framework, allowing it to encode continuous trajectories and sample future occupancy densities seamlessly over continuous time. Inference-time algorithms empower TrajFlow to generate probabilistic occupancy maps and infer likely future motion trajectories in a principled and tractable manner. Figure