PaperlayerView exact revision on arXiv

Source-linked AI summary

Naturalistic Driver Intention and Path Prediction using Recurrent Neural Networks

Alex Zyner, Stewart Worrall, Eduardo Nebot

arXiv:1807.09995v1cs.CV

TL;DR

Driver-intention prediction at unsignalized intersections remains difficult because vehicle movement is highly variable and multi-modal. The paper combines recurrent neural networks with a mixture density output and clustering to produce ranked trajectory possibilities with uncertainty, evaluating them on a large naturalistic dataset. The method outperformed all baselines on 5952 real-world trajectories.

  • Problem

    Driver intention at unsignalized intersections is difficult to predict because these scenes contain highly variable vehicle movement and interactions.

  • Method

    The method combines recurrent neural networks and a mixture density network with clustering that produces ranked possible trajectories with uncertainty.

  • Results

    The algorithm was tested on 5952 real-world trajectories and outperformed all baselines.

  • Takeaways & Limitations

    The paper provides a multi-modal probabilistic driver-intention and path-prediction method validated across five real-world roundabouts.

Abstract

from arXiv · show

Understanding the intentions of drivers at intersections is a critical component for autonomous vehicles. Urban intersections that do not have traffic signals are a common epicentre of highly variable vehicle movement and interactions. We present a method for predicting driver intent at urban intersections through multi-modal trajectory prediction with uncertainty. Our method is based on recurrent neural networks combined with a mixture density network output layer. To consolidate the multi-modal nature of the output probability distribution, we introduce a clustering algorithm that extracts the set of possible paths that exist in the prediction output, and ranks them according to likelihood. To verify the method's performance and generalizability, we present a real-world dataset that consists of over 23,000 vehicles traversing five different intersections, collected using a vehicle mounted Lidar based tracking system. An array of metrics is used to demonstrate the performance of the model against several baselines.

I. INTRODUCTION

The paper targets driver-intention prediction at dynamic, unsignalized urban intersections, where existing ADAS research has focused less and accident risk is substantial. It contributes a probabilistic, multi-modal prediction model, a naturalistic dataset, and demonstrations of limited-map operation and cross-intersection generalization.

  • Motivation: Unsignalized neighborhood intersections receive less attention than highways and structured signalized intersections despite their importance for autonomous-vehicle navigation.The paper notes that 84% of intersection accidents are attributed to driver recognition or decision error.
  • Contributions: The proposed model produces multi-modal path predictions with uncertainty and clusters them into a meaningful output.The contributions describe ranking possible paths to better represent predicted driver intention.
  • Dataset: The data is real-world and real-time, retaining tracking noise and vehicle-perspective issues because it was collected using a Lidar-enabled vehicle without filtering.This differs from overhead-camera datasets and avoids relying on future data during filtering or analysis.
  • Scope: The algorithm requires only knowledge that the vehicle is approaching a roundabout, rather than high-definition maps.The paper also demonstrates generalization between similar-sized intersections using data spanning multiple sites.

II. RELATED WORK

Prior work spans maneuver classification, single-trajectory prediction, and interaction-aware modeling, with much of the evidence centered on highways or structured and overhead-camera settings. The paper positions unsignalized intersection prediction as a comparatively less studied problem and contrasts real-world collection with limitations of common datasets and simulators.

  • Research categories: Driver-intention research is commonly grouped into maneuver-based, path-prediction, and interaction-aware models.These categories differ in whether they classify maneuvers, predict future trajectories, or incorporate surrounding-vehicle interactions.
  • Maneuver-based models: Maneuver-based models classify actions such as lane changes, stops, and intersection turns before associating trajectories with the predicted maneuver.Reported approaches include Support Vector Machines, Hidden Markov Models, Bayesian Networks, and LSTM-based classification.
  • Path prediction: Path-prediction models estimate future vehicle trajectories using physical models and methods such as Switching Kalman Filters, Monte Carlo Simulations, or Variational Gaussian Mixture Models.These methods generally produce a single prediction proposal with uncertainty, although maneuver recognition can enable multi-modal prediction.
  • Interaction-aware models: Interaction-aware models incorporate the movement or predicted movement of surrounding vehicles, with prior work emphasizing highway data from visual, radar, or simulated Lidar inputs.Related methods also use social-pooling or other mechanisms to model interactions in vehicle and pedestrian sequences.
  • Recurrent models: Recurrent neural networks have been used for maneuver classification and sequence generation, including autoregressive production of longer trajectories from predictive distributions.This establishes the sequence-modeling background for the paper’s RNN-based approach.
  • Dataset gap: Intersection research is less developed than highway research, while common datasets often rely on overhead cameras, processed tracking, or simulated environments.The paper identifies tracking noise and unrealistic collisions in NGSIM and difficulty validating simulator behavior for risky real-world scenarios.

III. PROBLEM DEFINITION

The paper targets driver-intention prediction at unsignalized urban intersections, where varied driving styles and nonlinear trajectories complicate standard tracking. It defines prediction from approach observations to future roundabout trajectories in a common coordinate frame.

  • Unsignalized intersections permit varied driving styles because they lack infrastructure enforcing strict vehicle behavior and ordering.
  • The dataset contains vehicle tracks with lateral and longitudinal positions, heading, and velocity, split into snippets for prediction.
  • Each input observation x_t contains position, velocity, and orientation, while Y represents the future track and Ŷ its probability estimate.
  • Roundabout maneuver lengths differ, with left turns shorter than right turns, so one fixed prediction horizon cannot represent every path.
  • Vehicle recordings are normalized to a common frame in which each vehicle enters from the bottom, with the origin near the intersection approach.
  • The model uses observations during the vehicle’s approach to predict its trajectory while traversing the roundabout.

IV. MODEL

The model uses a recurrent encoder-decoder network to process sequential vehicle observations and generate trajectory predictions. Normalization is integrated into the network and outputs are rescaled to real-world units.

  • The proposed model is a Recurrent Neural Network suited to sequential time-series inputs and variable-length outputs.
  • The architecture is an encoder-decoder with shared weights: the encoder processes observations, and the decoder generates the predicted path.
  • Unlike one-step feedforward approaches, this model generates a full prediction during training and uses recurrence to produce a whole track.
  • Normalization is implemented as the network’s first layer using fixed training-set parameters, and the final output is converted back to real-world units.

2) Sequence Length Padding:

The model handles maneuver-dependent sequence lengths with padding, allowing shorter trajectories to be represented without biasing later predictions toward longer paths.

  • Left-turn vehicles exit sooner than vehicles traveling straight or turning right, producing naturally shorter output sequences.
  • A padding logit lets the network identify when a vehicle has left the intersection and subsequent values are padding.
  • Padding repeats the vehicle’s last known position after exit, preventing later predictions from over-fitting to longer right-turn sequences.
  • Regression is unsuitable because averaging distinct plausible paths can produce a trajectory that is not itself a valid solution.
  • The Mixture Density Network represents multiple trajectory solutions through a weighted mixture of Gaussian distributions.
  • The output parameters include padding probability, mixture weights, two-dimensional means and standard deviations, and correlation coefficients.

4) Loss:

Training combines trajectory reconstruction likelihood with a padding cross-entropy objective, while feedforward sampling is examined because multimodality can cause the model to ignore sampled inputs.

  • The reconstruction loss measures the likelihood that ground truth could be sampled from the predicted mixture-density distribution.
  • Padding prediction is trained with cross entropy against the ground-truth padding label.
  • The final objective combines padding cross entropy with negative log likelihood, balanced by hyperparameters α and β.
  • During feedforward sampling, velocity and orientation are computed from the displacement between consecutive predicted positions.
  • Multimodal predictions can yield physically impossible sampled trajectories, causing the network to ignore feedforward inputs during prediction.

C. Output Consolidation

The model’s Gaussian-mixture outputs are consolidated into meaningful, likelihood-ranked paths using the Multi-PAC clustering algorithm. This produces a variable set of possible trajectories rather than a fixed number of solutions.

  • Multi-PAC consolidates Gaussian-mixture outputs into meaningful paths representing the model’s multiple modes.It first removes low-weight mixtures, then groups remaining outputs into clusters and constructs paths through time.
  • DBSCAN groups mixture centroids at each timestep, using probability-weighted averages to define cluster centroids.
  • A tree links cluster nodes across timesteps by assigning children to the closest parent, yielding one path from each leaf to the root.
  • The resulting paths are ranked by relative probability, providing a succinct representation of the multimodal output.
  • The clustering approach avoids restricting the algorithm to a predetermined number of solutions.

D. Proposed Models

The study evaluates three recurrent model variants that differ in how decoder inputs and losses are handled during sequence generation and training. These models are tested within the roundabout trajectory-prediction methodology.

  • The tested architectures comprise RNN-FF, RNN-ZF, and RNN-FL variants.
  • RNN-FF feeds a sampled output into the next timestep during training and prediction, while gradients propagate through recurrent layers but not the sampler.
  • RNN-ZF uses zero inputs for the next timestep during both path generation and training, with loss generated from the whole ground-truth sequence.
  • RNN-FL runs the decoder for one timestep and generates loss only for t+1, following an approach used in earlier mixture-density sequence tasks.
  • The methodology section covers algorithm testing, collected data, and comparison models.

1) Data Collection:

Data were collected from five Sydney roundabouts using a vehicle-mounted Lidar tracking system, with vehicle tracks labeled by entrance, exit, and maneuver. One complete intersection was held out for testing.

  • Data Collection: The tracking vehicle used six four-beam Lidars and an ibeo.HAD system for real-time road-user detection, classification, and tracking.Each Lidar has a 110-degree field of view, 200-metre range, and 25 Hz recording frequency.
  • Data Collection: The dataset covers five Sydney roundabouts selected for relatively high throughput, balanced traffic, and unobstructed visibility from the recording vehicle.
  • Data Collection: Only vehicles approaching from the recording vehicle’s road or directly opposite were used because those approaches provided more than 50 metres of visibility.
  • Data Collection: Tracks were labeled by entrance and exit as left, straight, right, or u-turn maneuvers, with u-turns excluded from results because they were exceedingly rare.
  • Model Training: Oliver-Wyndora was reserved as an entirely unseen test intersection to assess generalizability, while training used balanced destination classes through oversampling.

C. Metrics

The evaluation combines trajectory-distance metrics with baseline comparisons and selects both the highest-probability and most-accurate predicted paths from the multimodal output. Modified Hausdorff Distance addresses spatial divergence without penalizing temporal misalignment.

  • Evaluation uses one track snippet per vehicle, sampled when the vehicle crosses the intersection entrance, and compares multimodal predictions with several models.
  • Both the highest-probability path and the most-accurate path are evaluated to identify cases where the model entirely misses the ground truth.
  • Metrics: Total Euclidean error sums pointwise path errors across every timestep, while horizon Euclidean error compares predictions and ground truth at specified time horizons.The selected horizons are 1.2 seconds and 2.8 seconds.
  • Metrics: Modified Hausdorff Distance compares each track with the closest points on the other, penalizing spatial divergence without penalizing temporal misalignment.
  • Baselines: The metric set includes constant-velocity, constant-turn-rate-and-velocity, constant-turn-rate-and-acceleration, and Gaussian-process comparison models.

VI. RESULTS

The RNN-FF model’s most likely path outperformed all baselines, particularly for left- and right-turning vehicles, while its multimodal predictions could contain the correct path even when ranked lower.

  • The evaluation grouped results by vehicle destination to show performance across travel directions and address class imbalance across all tracks.
  • The tables report both the RNN-FF model’s most likely path and lowest-error path, revealing false negatives when the proper prediction was missed.
  • The RNN-FF model’s most likely path outperformed all baselines, especially for vehicles turning left or right.
  • The best-scoring RNN-FF path outperformed every baseline, indicating that the ground truth was contained in the multimodal predictions despite lower likelihood.

B. Qualitative Results

Qualitative results show that the model represents uncertainty over feasible maneuvers and becomes more confident as vehicles provide clearer evidence of their intentions, while difficult stops and cross-intersection shifts remain challenging.

  • Qualitative Results: The model’s white clustered paths expose multimodal uncertainty alongside observations, ground truth, and baseline predictions.
  • Qualitative Results: The model produces substantial uncertainty when a vehicle stops before entering the intersection, where intention prediction is especially difficult.
  • Qualitative Results: For a vehicle far from the intersection, the algorithm suggests three possible paths because all maneuvers remain feasible.
  • Qualitative Results: As a vehicle enters and travels through the intersection, left turns become infeasible and the model increasingly weights and then favors a right turn.
  • Ablative Results: The RNN-FL model matched the CTRA model’s worst results, supporting complete forward path prediction during training rather than loss only at the first time-step.
  • Ablative Results: Differences between RNN-ZF and RNN-FF were minimal, while single samples from multimodal distributions produced noisy outputs by mixing modalities.
  • Conclusion: The method was evaluated on 5,952 real-world trajectories and outperformed all baselines.
Loading 1807.09995v1…