Source-linked AI summary

Probabilistic Prediction of Vehicle Semantic Intention and Motion

Yeping Hu, Wei Zhan, Masayoshi Tomizuka

arXiv:1804.03629v1cs.LGstat.ML

TL;DR

Vehicle prediction needs to accommodate varied driving scenarios, uncertain behavior, and both destination and temporal information. SIMP addresses this with a probabilistic deep-neural-network framework based on semantic insertion areas, and the reported evaluations outperform SVM and QRF in prediction error and confidence intervals.

  • Problem

    Existing intention-prediction methods often fix the maneuver set for a specific scenario, while vehicle prediction also needs destination and timing information across varied traffic environments.

  • Method

    SIMP uses semantic-defined insertion areas and a probabilistic deep-neural-network framework to estimate vehicle intentions, final locations, and corresponding times.

  • Results

    SIMP outperforms SVM and QRF in prediction error and confidence intervals, with AUC values of 1 for Areas 1–3 and 0.994 for Area 4.

  • Takeaways & Limitations

    The predicted goal positions and times can be further used to generate optimal trajectories for predicted vehicles and a desirable path for the autonomous vehicle.

Abstract

from arXiv · show

Accurately predicting the possible behaviors of traffic participants is an essential capability for future autonomous vehicles. The majority of current researches fix the number of driving intentions by considering only a specific scenario. However, distinct driving environments usually contain various possible driving maneuvers. Therefore, a intention prediction method that can adapt to different traffic scenarios is needed. To further improve the overall vehicle prediction performance, motion information is usually incorporated with classified intentions. As suggested in some literature, the methods that directly predict possible goal locations can achieve better performance for long-term motion prediction than other approaches due to their automatic incorporation of environment constraints. Moreover, by obtaining the temporal information of the predicted destinations, the optimal trajectories for predicted vehicles as well as the desirable path for ego autonomous vehicle could be easily generated. In this paper, we propose a Semantic-based Intention and Motion Prediction (SIMP) method, which can be adapted to any driving scenarios by using semantic-defined vehicle behaviors. It utilizes a probabilistic framework based on deep neural network to estimate the intentions, final locations, and the corresponding time information for surrounding vehicles. An exemplar real-world scenario was used to implement and examine the proposed method.

I. INTRODUCTION

Vehicle prediction must handle diverse traffic scenarios and uncertain driver behavior, while combining semantic intentions with destination and timing information for useful motion forecasts.

  • I. INTRODUCTION: Existing intention classifiers are typically designed for one scenario with limited maneuvers, motivating methods that adapt across traffic environments.Generic maneuver categories may require manually selected thresholds and can fail when unclassified maneuvers occur.
  • I. INTRODUCTION: Semantic intentions represent entering insertion areas, allowing one unified formulation across varying road structures and driving scenarios.Insertion areas may be gaps between vehicles or lane entrances and exits.
  • I. INTRODUCTION: Predicting destinations can improve long-term motion forecasting by representing dynamics and automatically incorporating environmental constraints for unreachable regions.Destination prediction is presented as an alternative to directly predicting a single trajectory.
  • I. INTRODUCTION: Time-to-destination information complements goal locations and is important for subsequent ego-vehicle trajectory planning.Prior work directly predicts temporal quantities such as time-to-lane-change.
  • I. INTRODUCTION: SIMP uses a probabilistic deep-neural-network framework to estimate semantic intentions, final locations, and corresponding times while accounting for driver and traffic uncertainty.The paper defines its semantic question as which area the vehicle will most likely enter, where, and when.

II. CONCEPT OF SEMANTIC-BASED INTENTION AND MOTION PREDICTION (SIMP)

The method uses Mixture Density Networks to model conditional output distributions, with Gaussian mixture parameters constrained to form valid probability densities and trained by likelihood minimization.

  • II. CONCEPT OF SEMANTIC-BASED INTENTION AND MOTION PREDICTION (SIMP): A Mixture Density Network combines an artificial neural network with a mixture-density model to represent flexible conditional distributions p(y|x).The network maps input features to parameters of a mixture model over the output.
  • II. CONCEPT OF SEMANTIC-BASED INTENTION AND MOTION PREDICTION (SIMP): The Gaussian-mixture MDN maps inputs to mixing coefficients, means, and variances that define the output probability density.The mixing coefficients must be positive and sum to 1, while standard deviations must be positive.
  • II. CONCEPT OF SEMANTIC-BASED INTENTION AND MOTION PREDICTION (SIMP): MDN training minimizes negative log-likelihood over the training data and supports optimization through back propagation.The loss is defined over n training examples.

B. Proposed SIMP Method

SIMP assigns probability distributions to insertion areas, jointly modeling where and when a vehicle will enter each area. A DNN outputs area weights and 2D GMM parameters, trained to align area selection with the correct location and time.

  • B. Proposed SIMP Method: SIMP models each insertion area with a 2D Gaussian mixture over final insertion location and arrival time.The predicted vector is y = [y_s, y_t]^T, and multiple GMMs represent the possible areas in a scenario.
  • B. Proposed SIMP Method: The DNN outputs both the parameters of every area-specific 2D GMM and a weight w_a for each insertion area.The area weights represent the probability of entering the corresponding area.
  • B. Proposed SIMP Method: The loss combines area-weight learning with the likelihood of the correct location and time, using one-hot final-area labels.W1 and W2 are manually tuned so the two loss components have comparable magnitudes during training.
  • B. Proposed SIMP Method: The SIMP architecture uses current-step features under a first-order Markov assumption, fully connected layers, dropout, and constrained output transformations.For N_a insertion areas, the network produces N_a ∗ (M ∗ 6 + 1) output parameters.

III. AN EXEMPLAR HIGHWAY SCENARIO

The paper applies SIMP to an exemplar highway scenario and presents its data source and problem formulation before evaluation.

  • III. AN EXEMPLAR HIGHWAY SCENARIO: An exemplar highway scenario is used to apply SIMP, with the data source and problem formulation presented for this application.

A. Dataset

The exemplar scenario uses NGSIM US 101 highway trajectories and represents possible vehicle outcomes through dynamic insertion areas. These areas distinguish lane changes from lane keeping while encoding location and timing.

  • A. Dataset: NGSIM US 101 provides highway vehicle trajectories sampled at 10 Hz from a measurement area about 640 meters long with five freeway lanes and an auxiliary lane.For lane-changing vehicles, up to 40 frames before lane-mark intersection are selected.
  • A. Dataset: The scenario selects a predicted vehicle, nearby reference vehicles, and surrounding vehicles, considering at most three lanes and seven vehicles per input frame.Vehicles judged too far from the predicted vehicle are treated as nonexistent within the scenario range.
  • A. Dataset: Five Dynamic Insertion Areas represent possible destinations; areas 1–4 indicate lane changes, while area 5 indicates lane keeping.The areas are dynamic because their locations and sizes vary at each time step.
  • A. Dataset: In the highway setting, y_s is the distance from the final insertion point to its reference vehicle and y_t is time-to-lane-change.TTLC equals zero at lane-mark intersection and is set to 4 seconds for lane keeping.

C. Features and Structure Details

Each input frame uses vehicle and roadway features together with missing-vehicle conventions, and the exemplar network uses three fully connected layers, dropout, and five insertion areas.

  • C. Features and Structure Details: Each input frame contains 25 selected features and labels for area weight, final goal location, and remaining insertion time.The feature design uses roadway geometry and relative vehicle information.
  • C. Features and Structure Details: Absent vehicles are encoded with large longitudinal distance and matched velocity, while nonexistent lanes receive fixed proxy vehicle configurations.These conventions preserve a fixed-size input representation.
  • C. Features and Structure Details: The network has three fully connected 400-neuron layers with tanh activations, followed by dropout at rate 0.5.The exemplar scenario sets N_a to five.

A. Evaluation Setup

The evaluation compares SIMP with SVM for intention prediction and uses ROC-based threshold selection, precision, recall, F1, and prediction-time analyses.

  • SIMP is evaluated against SVM for intention prediction using simplified three-class labels and ROC curves.The evaluation includes lane-change and lane-keeping intention classes.
  • SIMP outputs probabilistic predictions through a framework evaluated alongside QRF and SVM baselines.The supplied setup passage identifies SVM as the intention-classification comparator.
  • Example visualizations sample 50 points per testing frame by allocating samples according to DIA weights and area-specific distributions.The horizontal axis is measured in feet.
  • The evaluation selects a ROC-derived threshold, then measures recall, precision, F1 score, and average prediction time.The later DIA analysis uses the original five semantic intention classes.

3) Evaluation for Motion Prediction:

Motion prediction evaluates probabilistic destination and TTLC outputs using quantile intervals for uncertainty assessment.

  • The two motion outputs are final locations in insertion areas and remaining time to reach those locations, modeled with separate quantile regressors.The labels are either location or time information.
  • Testing evaluates each motion method with 68% and 95% quantile intervals.QRF-68% spans the 16th–84th percentiles, while QRF-95% spans the 2.5th–97.5th percentiles.

B. Results and Discussion

The selected cases show plausible DIA assignments and continuity, while SIMP outperforms SVM in intention accuracy and prediction timing.

  • 1) Visualization of Selected Cases:: The lane-change example samples points in the proper DIA for each frame, with multiple possible areas appearing early before a specific choice is made.The passage notes that numerical correctness is difficult to establish before human-labeled ground truth is available.
  • 1) Visualization of Selected Cases:: When a reference vehicle switches during driving, sampled points remain in the correct DIA and tightly distributed around the target line despite a change between frames 19 and 20.The case tests whether predictions avoid large discontinuities during changing reference-vehicle conditions.
  • 2) Intention Estimation:: SIMP outperforms SVM in both prediction accuracy and average prediction time, according to the overall comparison.The evaluation uses ROC-derived precision, recall, and F1 measures alongside prediction timing.
  • 2) Intention Estimation:: SIMP forecasts lane-change intention approximately 2s before the actual lane change and achieves AUC values of 1 for Areas 1–3 and 0.994 for Area 4.These results indicate accurate detection of both lane-change direction and the specific DIA in the selected time window.

3) Motion Prediction:

Compared with QRF, SIMP produces lower errors for TTLC and destination prediction, with especially concentrated uncertainty for TTLC as lane change approaches.

  • Fig. 5 visualizes TTLC means and 1σ and 3σ prediction intervals from 100 samples for each related DIA.Areas with weights too small to receive sampled points are shown in gray.
  • SIMP has lower RMSE than QRF for both TTLC and destination prediction, with TTLC error tending toward zero in lane-change cases.Destination error remains nonzero at t = 0 because the prediction is relative to a reference vehicle whose velocity may vary.
  • TTLC prediction intervals progressively contract toward the ground truth as TTLC approaches zero.The decreasing one- and two-standard-deviation intervals indicate a more centralized predicted Gaussian distribution.

V. CONCLUSIONS

SIMP combines semantic intention and motion prediction in a probabilistic framework that generalizes across traffic scenarios and outperforms SVM and QRF on prediction error and confidence intervals. Its predicted goal positions and times support trajectory generation, while more complex scenarios and vehicle occlusion remain future-work targets.

  • SIMP outperforms SVM and QRF in prediction error and confidence intervals for intention and motion prediction.
  • The framework combines prediction tasks through semantics to generalize across different traffic scenarios.
  • Predicted goal positions and time information can support optimal trajectories for surrounding vehicles and a desirable ego-vehicle path.
  • Future work will examine SIMP in more complex scenarios and account for vehicle occlusion.
Loading 1804.03629v1…