Evaluating Generative Vehicle Trajectory Models for Traffic Intersection Dynamics

Yash Ranjan, Rahul Sengupta, Anand Rangarajan, Sanjay Ranka · 2025

Read Evaluating Generative Vehicle Trajectory Models for Traffic Intersection Dynamics by Yash Ranjan, Rahul Sengupta, Anand Rangarajan, Sanjay Ranka (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-2506-08963.txt.

Introduction

Yash Ranjan, Rahul Sengupta, Anand Rangarajan, and Sanjay Ranka wrote a 2025 paper on how to judge generative models of vehicle motion at traffic intersections. Intersections guide the flow of people and goods, but paths cross there and crashes are common.

Most tests of such models use reconstruction errors alone and skip live use. The authors built a tool to train, run, and score models in a live microsimulation, with traffic measures such as red-light violations and unallowed stops. They trained a top multi-vehicle forecasting model on data from a calibrated scene of a real urban site.

Catalog record

CREATORYash Ranjan, Rahul Sengupta, Anand Rangarajan, Sanjay Ranka
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE26 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-08963

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Evaluating Generative Vehicle Trajectory Models for Traffic Intersection Dynamics Yash Ranjan, Rahul Sengupta, Anand Rangarajan, Sanjay Ranka Abstract Traffic Intersections are vital to urban road networks as they regulate the movement of people and goods. However, they are regions of conflicting trajectories and are prone to accidents. Deep Generative models of traffic dynamics at signalized intersections can greatly help traffic authorities better understand the efficiency and safety aspects. At present, models are evaluated on computational metrics that primarily look at trajectory reconstruction errors. They are not evaluated online in a ‘live’ microsimulation scenario. Further, these metrics do not adequately consider traffic engineering-specific concerns such as red-light violations, unallowed stoppage, etc. In this work, we provide a comprehensive analytics tool to train, run, and evaluate models with metrics that give better insights into model performance from a traffic engineering point of view. We train a state-of-the-art multi-vehicle trajectory forecasting model on a large dataset collected by running a calibrated scenario of a real-world urban intersection. We then evaluate the performance of the prediction models, online in a microsimulator, under unseen traffic conditions. We show that despite using ideally-behaved trajectories as input, and achieving low trajectory reconstruction errors, the generated trajectories show behaviors that break traffic rules. We introduce new metrics to evaluate such undesired behaviors and present our results. 1 Introduction Traffic intersections play a critical role in the functioning of urban road networks. They ensure the smooth flow of vehicles and pedestrians by ensuring their trajectories do not conflict, avoiding congestion and accidents. Hence, dynamics seen at signalized urban intersections are complex, as various road users (vehicles and pedestrians) must not only interact with each other, but they must also pay attention to the traffic signal state as well as road markings indicating lanes and crosswalks. Computational-modeling of such scenarios was usually done on a case-by-case basis using agent-based micro-simulators, which can be computationally-heavy. Intelligent Transportation Systems (ITS) has revolutionized the way sensor data from urban intersections is acquired, processed, and utilized. Various sensor modalities such as inductive loop detectors, signal state recorders, video cameras, LiDAR, radar etc. provide comprehensive high-resolution data. Advances in Cloud Storage technologies have provided scalable data storage for vast amounts of multimodal sensor data. High-performance computing hardware (such as GPUs/TPUs) has accelerated the development of Deep Learning models trained on large datasets. Hence, there has been a great deal of published literature on the application of Deep Learning models to ITS. One important area of application is the trajectory prediction of road users. Usually, trajectory data is generally obtained by performing object-tracking on video/LiDAR data. This often requires large data processing resources as well as advanced vision-processing algorithms [ 5 ] . 1.1 Related Work There are many works that deal with trajectory prediction for road users. Important among them include Trajectron (and its successor Trajectron++) [ 15 ] , which represents a traffic scene as a directed spatio-temporal graph and produces dynamically-feasible trajectories using a graph-structured recurrent model. Another important model is TNT (target-driven trajectory) [ 19 ] which consists of three stages that are trained end-to-end: the first predicts an agent’s potential target states, the second generates trajectory state sequences conditioned on targets and the final stage estimates trajectory likelihoods and selects a final compact set of trajectory predictions. Surveys such as [ 8 ] , [ 10 ] , and [ 1 ] discuss various methods for vehicle trajectory prediction. However, they focus on autonomous vehicles and do not specifically deal with the modeling of intersection behaviors. [ 6 ] surveys commonly used performance metrics and robustness strategies for vehicle trajectory prediction. [ 9 ] discusses commonly-used metrics and their drawbacks in the context of multimodal trajectory prediction. [ 3 ] introduces new metrics for trajectory prediction apart from commonly-used distance-based metrics. However, it does not focus on intersection-related behaviors, especially the effect of traffic signals and compliance thereof. 1.2 Our Contribution While there is a good deal of relevant literature on deep generative models for trajectory prediction, there is a lack of metrics focusing on intersection dynamics in a multi-vehicle scenario. Present metrics are written from the point-of-view of autonomous vehicles, where the prediction of road users’ trajectories revolves around the perception (via sensors) of the ego vehicle. This is not helpful to road traffic authorities who study the intersection dynamics as a whole, with trajectory data from intersection-mounted video cameras, LiDAR, DSRC/CV2X, drone-based cameras etc. These provide a birds-eye view of multiple vehicles at the intersection and their movements. Further, the evaluation of these models is generally done using a pre-recorded test set. It is not possible to generate new arrival patterns or change other parameters, such as signal timing plan. Given the critical importance of signalized intersections to road traffic authorities, our work makes the following contributions: 1. We develop an end-to-end framework for evaluating and benchmarking Deep Generative models for the task of multi-vehicle trajectory prediction at a signalized intersection. Given the trajectory prediction model architecture code, the system will train the model on the testbed intersection and vehicle flows, then evaluate the trained model and report its performance. 2. We introduce new evaluation metrics that address traffic-specific concerns such as traffic signal compliance, unnecessary stoppage, and braking behavior. These are novel metrics not usually seen in published trajectory prediction literature. 3. We include intersection-based vehicle position encoding and signal timing information in our trajectory prediction model, which is not usually seen in published literature. 4. We test the models using a simulation-in-loop real-world signalized urban intersection with calibrated vehicle flows. 2 Trajectory Prediction Model for Signalized Intersection Scenario Our problem requires us to model the spatial-temporal relationship between an object’s history with the state of the objects around it to predict the future position. The model used is largely based on [ 18 ] with key changes to incorporate intersection-specific details. We discuss the different components of the model and how we incorporate the external factors which are the signal, intersection geometry, and the object position. Figure 1: The model architecture. The data comes from the simulator and is passed through the model to get the distribution over the future state. The mode of the distribution is sampled and sent back to the simulator, which visualizes it. 2.1 Trajectory Prediction The objective of the trajectory prediction model is to generate plausible trajectory distributions for different agents (such as vehicles and pedestrians) that interact in a scene. This time-dependent number of agents is denoted by n(t) , representing the interacting agents A_{1},\ldots,A_{n(t)} . In this work, we focus only on vehicles. Each agent considers its own motion history, relevant traffic signals, potential interactions with neighboring agents, and its position relative to ‘static points’ (described later). The historical data for agent A_{i} at time t is expressed as X^{t-H+1,\ldots,t}_{i} , where H represents the number of history steps. The traffic signal for the route class c of agent A_{i} is denoted as s^{t}_{c} , its neighboring agents as N^{t}_{i} , and relative position R^{t}_{i} . The model outputs a probability distribution over the future trajectory Y^{t+1,\ldots,t+T}_{i} for the next T steps, formulated as p(Y^{t+1,\ldots,t+T}_{i}|X^{t-H+1,\ldots,t}_{i},N^{t}_{i},R^{t}_{i},s^{t}_{c}) . Here X consists of rectangular coordinates (x,y) , velocity (\dot{x},\dot{y}) , acceleration (\ddot{x},\ddot{y}) and

Read the complete text