Probabilistic Risk Metrics for Navigating Occluded Intersections

Abstract

Among traffic accidents in the US, 23% of fatal and 32% of non-fatal incidents occurred at intersections. For driver assistance systems, intersection navigation remains a difficult problem that is critically important to increasing driver safety. In this paper, we examine how to navigate an unsignalized intersection safely under occlusions and faulty perception. We propose a real-time, probabilistic, risk assessment for parallel autonomy control applications for occluded intersection scenarios. The algorithms are implemented on real hardware and are deployed in a variety of turning and merging topologies. We show phenomena that establish go/no-go decisions, augment acceleration through an intersection and encourage nudging behaviors toward intersections. Index Terms-Intelligent Transportation Systems; Human Factors and Human-in-the-Loop; Autonomous Vehicle Navigation I. INTRODUCTION I NTERSECTIONS present one of the most challenging driving scenarios because a vehicle must interact with others to navigate safely. In 2016, 23% of fatal and 32% of non-fatal traffic incidents in the U.S. occurred at intersections 1. Recent advances in robotic perception and control routines promise to enhance the safety of passengers on the road through fully autonomous cars or by augmenting the human with advanced driver assistance systems (ADAS). However, reasoning about intersections remains a major challenge. Perception systems may fail to detect other vehicles as a consequence of occluded views. Causes of these occlusions may stem from cross-traffic behavior, buildings or road geometries that cause poor visibility [2]. As a consequence, tracking and detecting other road objects may fail in unpredictable ways. New models must account for the resulting uncertainty and risk while ensuring robustness to imperfect perception and remaining computationally efficient.

Limited Visibility Aware Motion Planning for Autonomous Valet Parking Using Reachable Set Estimation

“…Many works have addressed different aspects of risk assessment and safe planning for conditions of limited visibility. Some researchers have studied the probabilistic risk assessment of occluded areas [ 12 , 13 , 14 , 15 ]. One of them [ 12 ] proposed a probabilistic risk assessment method using the volume of traffic on the road, whereas [ 13 ] addressed a similar approach using a damage model based on the masses and velocities of two vehicles.…”

“…In addition, ref. [ 15 ] proposed a graphical model capable of describing the risk of a road segment over time, and then addressed the occupancy estimation using a dynamic Bayesian network. Even though these studies explicitly represent the risk as a probability such that the risk can be considered in the planning process, expressing the risk as a probability cannot ensure a provable safety.…”

Reasoning about Potential Hidden Traffic Participants by Tracking Occluded Areas

“…Another direction of research focuses on estimating collision risk caused by the potential hidden TPs in occlusions, e.g. [7], [8] and [9]. [7] obtains a comfort plan by utilizing an optimization-based motion planner to reduce the risk, while [8] evaluates the driver's current behavior in terms of risk, warns the driver in case of critical events.…”

“…[7], [8] and [9]. [7] obtains a comfort plan by utilizing an optimization-based motion planner to reduce the risk, while [8] evaluates the driver's current behavior in terms of risk, warns the driver in case of critical events. As they regard risk as soft constraints, safety guarantee is not provided.…”

Prediction of Pedestrian Risky Level for Intelligent Vehicles

“…Since predefined dynamic models may fail to account for the movement uncertainty, several probabilistic collision prediction methods were proposed. For example, the work [6] quantified risk using a probabilistic graphical model and considering the likelihoods at all potential intersections. The risk to the ego vehicle was computed as the expected number of incidents had the ego vehicle entered the intersection at a specific time.…”