Technical Summary: A Hierarchical Pedestrian Behavior Model for Traffic Simulation

Problem Statement

The development and testing of autonomous vehicles (AVs) and automated driving systems (ADS) require rigorous scenario-based testing to ensure safe interactions with pedestrians, the most vulnerable road users. Existing microscopic pedestrian models in traffic simulation often suffer from significant limitations:

• Lack of Extensibility: Many models focus heavily on specific interaction scenarios, making them difficult to adapt to new contexts.
• Black-Box Nature: Approaches using non-deterministic or "black-box" methods (e.g., certain data-driven or cellular automata models) do not allow engineers to derive or control specific pedestrian actions based on scenario inputs.
• Inflexibility: While some models demonstrate higher-level behaviors (e.g., lane formation), they generally lack the ability for engineers to customize behaviors or inject specific, desired actions (such as rare or critical maneuvers) into a test scenario.
• Trajectory Rigidity: Simple trajectory-based models that follow pre-defined speed profiles cannot dynamically adjust to ensure specific outcomes (e.g., a guaranteed collision for safety testing) without constant manual reconfiguration.

There is a need for a pedestrian simulation model that is both highly controllable (allowing explicit injection of behaviors for testing) and realistic (replicating natural human decision-making and motion).

Methodology

The authors propose a hierarchical pedestrian behavior model integrated into the GeoScenario Server, a scenario definition and execution engine. The model is structured into three distinct layers:

1. Behavior Layer (High-Level Decision Making)

• Mechanism: Utilizes Behavior Trees to model decision-making processes explicitly.
• Function: At each simulation cycle, the behavior tree is "ticked" to select an appropriate maneuver based on the current environmental state (e.g., traffic light status, proximity of vehicles, destination).
• Customizability: Each pedestrian is assigned a behavior tree file. The model includes a library of reusable trees and a catalog of maneuvers (e.g., Enter Crosswalk, Wait at Crosswalk, Increase Walking Speed) and conditions (e.g., Crossing light green?, Vehicle approaching).
• Aggressiveness Levels: The model supports varying levels of pedestrian aggressiveness (Low, Medium, High) by modifying sub-trees. For instance, a "High" aggressiveness pedestrian may enter a crosswalk on a yellow or red light if safe, whereas a "Low" aggressiveness pedestrian waits for green.

2. Maneuver Layer (Trajectory Planning)

• Function: Translates the high-level maneuver selected by the Behavior layer into low-level instructions.
• Output: Generates a vector containing three components:
  1. A waypoint (intermediate goal).
  2. A desired direction unit vector.
  3. A desired speed.
• Role: This layer bridges the gap between abstract decisions and physical motion, ensuring the pedestrian's path aligns with the selected maneuver (e.g., directing a pedestrian to the end of a crosswalk).

3. Motion Planner Layer (Low-Level Motion Execution)

• Mechanism: Employs an adapted Social Force Model (SFM).
• Function: Calculates the pedestrian's acceleration and velocity changes based on physical forces:
  • Attracting Force: Pulls the pedestrian toward the destination or waypoint.
  • Repelling Forces: Pushes the pedestrian away from other pedestrians, vehicles, and environmental boundaries (walls/borders).
• Integration: The layer receives the trajectory instructions from the Maneuver layer and the real-time state of the environment (positions/velocities of other agents). It iterates the SFM equation to update the pedestrian's state (position, velocity, acceleration, heading) for the next time step.

Implementation

The model is implemented in Python and integrated into the GeoScenario Server. It supports the Lanelet2 map format for road structure representation. The system allows for the creation of "Evaluation Scenarios" where a simulated pedestrian (SP) replaces an empirical pedestrian (EP) from a real-world dataset to test interactions with other agents (vehicles and pedestrians) following their recorded trajectories.

Key Contributions

The paper outlines two primary contributions:

1. A Novel Hierarchical Model: A pedestrian motion simulation framework featuring:
   • Highly Controllable Decision Making: Achieved via Behavior Trees, providing a domain-specific language for engineers to customize pedestrian logic and inject specific behaviors (including rare/critical situations).
   • Realistic Motion: Achieved via an adapted Social Force Model that handles low-level trajectory execution and dynamic interactions.
2. Comprehensive Evaluation: An assessment of the model's performance on two real-world datasets from Ontario, Canada, with different road structures (a busy four-way intersection and a single unsignalized crosswalk).

Results

The model was evaluated on two metrics: Low-Level Trajectory Fidelity and High-Level Decision Accuracy.

Trajectory Fidelity

The model's ability to replicate real-world pedestrian paths was measured using Euclidean Distance (ED) and Fréchet Distance (FD).

• Intersection Dataset: The model achieved an average ED of 1.55 m (segmented scenarios) and 2.23 m (full scenarios), significantly outperforming a baseline trajectory approximation method (5.45 m).
• Single Crosswalk Dataset: The model achieved an average ED of 1.36 m.
• Observation: While the model sometimes deviated geometrically from the baseline due to reactive interactions (e.g., avoiding oncoming pedestrians on the opposite side of the path compared to the empirical data), it successfully captured the dynamic nature of pedestrian movement.

Decision-Making Accuracy

The model's ability to replicate high-level decisions was evaluated on the intersection dataset:

• Decision 1 (Crosswalk Selection): The model selected the same crosswalk as the empirical pedestrian 100% of the time (253/253 instances).
• Decision 2 (Crossing on Red/Yellow): The model matched the empirical decision 98.0% of the time (145/148 instances). The three mismatches occurred when empirical pedestrians waited for a red signal and then crossed immediately upon a specific traffic light change (two-way advanced green), a nuance that can be addressed by adjusting the "aggressiveness" parameters.

Critical Scenario Generation

The paper demonstrates the model's utility in generating a guaranteed collision scenario between a pedestrian and a vehicle at an unsignalized crosswalk. Unlike static trajectory models, the hierarchical model allows the pedestrian to dynamically adjust speed based on the vehicle's approach to ensure a collision occurs, regardless of initial positions or velocities. This capability is crucial for testing AV responses to critical situations without manual trial-and-error.

Significance and Claims

The authors claim that their work provides a valuable tool for scenario-based testing of AVs by offering a level of control and explicitness previously lacking in pedestrian simulation literature.

• Explicit Representation: By using Behavior Trees, the decision processes are transparent, allowing engineers to inject desired behaviors to test AV responses to critical or dangerous situations safely.
• Fidelity and Accuracy: The model replicates real-world pedestrian trajectories with high fidelity (average deviation of 1.36 m) and decision-making accuracy (98% or better) using only high-level routing information.
• Extensibility: The modular architecture allows for the easy addition of new maneuvers, conditions, and sub-trees to cover a wide range of conceivable behaviors, including those not present in existing datasets.

The paper concludes that this hierarchical approach effectively balances the need for controllability (for testing specific scenarios) with realism (for generating naturalistic human behavior), making it a viable solution for the rigorous testing required by autonomous driving development.