Autonomous driving behaviors for vulnerable road users

This is the article written by Automotive Intelligence team in KPMG Consulting.

The article presents a vision for the interaction of pedestrians and cyclists with Automated Driving System-Dedicated Vehicles (ADS-DVs), offering guidance on responding to the programmed behaviors of ADS-DVs in preparation for a new transportation system.

1. Definition of vulnerable road users (VRUs)
2. Why are interactions with VRUs difficult for ADS-DVs?
3. How should ADS-DVs react to VRUs?
4. What kind of operation should VRUs expect from ADS-DVs?

Vulnerable Road Users (VRUs) mainly include pedestrians and pedal cyclists, who are at higher risk of being injured or killed in crashes with vehicles compared to passengers secured and protected inside vehicles by passive safety systems such as airbags.

For ADS-DVs, interactions with such VRUs are important to consider for their development and deployment. VRUs are diverse in their appearance, kinematics, and behaviors. They could stop or move, for example, when riding a kick scooter. Road workers as VRUs or someone changing a tire are on or near the roadway affect traffic. Pedal cyclists may move with traffic, against it, or across it, while wheelchairs cross at an intersection. VRUs might follow traffic rules or choose not to follow them if they think no vehicle is approaching.

It can be challenging to detect VRUs, even from a reasonable distance, due to their physical properties and the challenge in predicting the future state or actions. The consequences of a detection or prediction error could be severe. Therefore, ADS-DVs should always consider possible errors and have a margin to correct them at any time.

Object and event detection and response (OEDR) is defined in the document of SAE (J3016)*¹ as consisting of monitoring of the driving environment (detecting, recognizing, and classifying objects and events, while preparing to respond as needed) in real-time and executing an appropriate vehicle response to the objects and events. OEDR is a fundamental element of executing the driving task of ADS-DVs and some external variables may change multiple times in less than a second. For example, VRUs may pose detection, classification, and response challenges, which could impact the path plan of ADS-DVs.

Challenges for detection include variability in shape, texture, color, material, and obscuration. Physical properties, including shapes, can be large, small, tall, or short. They could also vary in skin color, number of appendages, or type of locomotion. Shape outlines can vary or be obscured if a pedestrian carries something such as grocery bags or is covered with an umbrella. Clothing types are different by season and location. Color and texture can even blend in with the environment. Orientation or pose may vary, for eample, between a pedestrian and a person with mobility assistance.

VRU behaviors can be influenced by factors such as weather, environmental conditions, the presence of construction, and traffic control devices known to ADS-DVs. Additionally, VRU behaviors can also be influenced by internal factors, which cannot be perceived externally. VRUs might change heading more rapidly than vehicles do. For example, pedestrians could change direction in the middle of a crosswalk and move back to a sidewalk.

Object classification will improve prediction for ADS-DVs by narrowing the set of an object’s likely behaviors in a scenario, but like other object classes, prediction errors can arise from unknowable factors influencing VRUs’ decisions internally made. VRUs are possible to behave unpredictably if they are unfamiliar with local rules or in a group. For example, tourists in unfamiliar areas or pedestrians interacting with new types of traffic control devices may increase uncertainty.

Developers should design and test the end-to-end system level performance of their ADS-DVs to ensure safer driving tasks and behaviors in the presence of reasonably expected variations of VRUs in differing conditions.

ADS-DVs should have safety performance thresholds based on human driving behaviors, however, not all human driving behaviors are recommended.

Naturalistic driving studies are a method used in the transportation research to observe and analyze the real-world driving behaviors, such as collecting data from over 3,400 drivers and vehicles, recording more than 5.4 million trips and 36,000 crash-related events^*1. This approach allows researchers to collect data on how human drivers reasonably interact with their vehicles on the roads in natural, and everyday conditions, without the influence of experimental settings.

Like human drivers who interact with VRUs, ADS-DVs predict the potential positions of objects or available free space by making assumptions about objects, including their capacity to move. Manufacturers define a range of values for VRU kinematics such as heading angle, rate change of that, velocity, braking ability, and other factors. Data sources such as naturalistic driving data should be valuable to inform these values. Additional assumptions may also be tested for reasonable combinations of values (e.g., rate change of heading angle combined with velocity) to help define reasonably expected object behaviors.

ADS-DV developers should design the ADS-DVs to follow traffic rules. In this way, VRUs are not expected to make special considerations in the presence of the ADS-DVs. If an ADS-DV detects an object, but the object is not or cannot be classified, the ADS-DV can still plan a trajectory to reduce the probability of a conflict. For example, when an ADS-DV encounters an unclassified object, it plans more conservative maneuvers (such as increasing lateral or longitudinal separation) to protect the unknown object, which could be a VRU.

However, in many cases, overly conservative behaviors could cause confusion and frustration in other vehicles on the roads. To avoid this, if a detected object can be classified with sufficient confidence and ADS-DVs can make reasonable assumptions about the object’s class and environment, they may refine their behavior to improve mobility. For example, a detected object may be classified as a pedal cyclist legally riding in a shared lane. Using this classification, for the object identified as a pedal cyclist, ADS-DVs apply the appropriate safety margins defined in traffic rules and account for other contextual elements such as the environment and other road users to determine the next course of action. The motion planning considerations may result in the ADS-DVs‘ maintaining the speed and passing the cyclist or remaining behind the cyclist.

With a detected object now confidently classified, ADS-DVs can plan a refined set of maneuvers. These maneuvers factor in assumptions about the VRU class and the context (applicable traffic rules, environment, other road users, etc.) of the situation and determine how the risk will be managed.

VRUs should expect ADS-DVs to follow traffic rules. Compliance with traffic rules is an important safety metric for ADS-DVs. In cases where the law either requires or implies a judgment or a subjective assessment by a human driver, the ADS-DVs are expected to behave in a manner consistent with the traffic and conditions in which they were trained to operate.

VRUs and human-operated vehicles might not always follow the law and sometimes could decide to prioritize mobility or speed over safety. Past experiences, which were not always safe, or false consensus bias could lead some people to expect similar behaviors from other drivers (possibly both humans and ADS-DVs). ADS-DVs, which always follow traffic rules, could potentially cause conflicts with VRUs or other road users, who are possible to individually prioritize mobility over safety. However, as mentioned earlier, VRUs expect ADS-DVs to follow traffic rules for their own safety, thereby establishing it as the right consensus.

To make VRUs feel comfortable in the society with ADS-DVs, the ADS-DVs should be reliable enough on the roads so that the VRUs can believe that the ADS-DVs comply with traffic rules to protect the VRUs in any condition.

*1: SAE International ”Taxonomy and Definitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicles J3016_202104”

*Reference for Figure1: SAE Industry Technologies Consortia (Automated Vehicle Safety Consortium) “AVSC Best Practice for Interactions Between ADS-DVs and Vulnerable Road Users (VRUs)”

KPMG Consulting
Manager, Yukie Koyano