Intention-Aware Risk Estimation: Field Results Stéphanie Lefèvre, Dizan Vasquez, Christian Laugier, Javier Ibañez-Guzmán Abstract— This paper tackles the risk estimation problem from a new perspective: a framework is proposed for reasoning about traffic situations and collision risk at a semantic level, while classic approaches typically reason at a trajectory level. Risk is assessed by estimating the intentions of drivers and detecting conflicts between them, rather than by predicting the future trajectories of the vehicles and detecting collisions between them. More specifically, dangerous situations are iden- tified by comparing what drivers intend to do with what they are expected to do according to the traffic rules. The reasoning is performed in a probabilistic manner, in order to take into account sensor uncertainties and interpretation ambiguities. This framework can in theory be applied to any type of traffic situation; here we present its application to road intersections. The approach was validated with field trials using passenger ve- hicles equipped with Vehicle-to-Vehicle wireless communication modems. The results demonstrate that the algorithm is able to detect dangerous situations early and complies with real-time constraints. I. I NTRODUCTION Active safety systems are increasingly present in commer- cial vehicles, as part of a global effort to make roads safer. The purpose of such systems is to avoid or mitigate accidents through driver warnings or direct actions on the commands of the vehicles (braking, steering). At an algorithmic level, ac- tive safety functions rely on four processing steps: detect and track relevant entities in the environment (object assessment step), establish the relationship between these entities for a better understanding of the situation (situation assessment step), estimate the level of danger of the current situation (risk assessment step), and decide on the best course of action in order to promote safety (decision making step) [1]. The contribution of this paper concerns the second step. While classic approaches evaluate the risk of a traffic sit- uation by predicting the future trajectories of vehicles and detecting collisions between them, we propose to reason on risk at a semantic level. In the proposed framework, risk is assessed by estimating the intentions of drivers and detecting conflicts between them. Traffic rules are explicitly represented in our model, which makes it possible to estimate both what drivers intend to do and what they are expected to do. Conflicts are then identified by comparing intentions and expectations. This formulation of risk reflects the fact that most road accidents are caused by driver error [2], and has the advantage that it does not require to predict the future trajectories of vehicles. S. Lefèvre, D. Vasquez, and C. Laugier are with Inria Grenoble Rhône-Alpes, 655 av. de l’Europe - Montbonnot, 38334 Saint Ismier Cedex, France, {stephanie.lefevre, alejandro-dizan.vasquez-govea, chris- tian.laugier}@inria.fr. J. Ibañez-Guzmán is with Renault S.A.S., 1 av. du Golf, 78288 Guyancourt, France, javier.ibanez-guzman@renault.com The remainder of this paper is organized as follows. Section II reviews related work. Section III describes the proposed approach for risk assessment for general traffic sit- uations. Section IV describes the application of this general framework to road intersections. Section V presents results obtained in field trials with passenger vehicles negotiating a T-shaped give-way intersection. II. RELATED WORK By far the most popular approaches to risk estimation are based on trajectory prediction [3]. The idea is to use a motion model to predict the possible future trajectories of each vehicle in the scene, and then to look for intersections between pairs of trajectories to detect future collisions. Extensive research has been conducted on trajectory pre- diction. The most common solution is to rely on purely physical models (dynamic or kinematic) of the motion of a vehicle [4], [5], [6], [7], [8], [9], however those cannot reason at a high level about the situation and therefore are limited to short-term collision prediction. Long-term prediction can be improved by reasoning at a maneuver level instead of a purely physical level, and by taking into account the constraints imposed by the road network on the motion of vehicles. One strategy is to first identify the maneuver intention of each driver and to then generate trajectories corresponding to that maneuver intention on the current road layout. For this purpose, Aoude et. al. [10] used a combination of Support Vector Machines and Bayesian Filtering for maneuver intention estimation, and Rapidly-exploring Random Trees for trajectory generation. As an alternative, Laugier et. al. [11] proposed to combine Hidden Markov Models and Gaussian Processes. Another solution is to use Monte-Carlo simulation to explore the different possible realizations of a maneuver by sampling on the input variables [12]. Other works propose integrated frameworks based on Markov State Space Models (MSSM) which explicitly represent the maneuver intention of drivers and allow the prediction of future trajectories without the need of a separate “maneuver intention estimation” step [13], [14], [15]. The main limitation of approaches based on trajectory prediction is their high computational cost. For advanced motion models, which take into account the local shape of the road layout and the dependencies between the motion of different vehicles, the cost of computing all the possible future trajectories and the probability that they intersect is not compatible with real-time safety applications. The classic solution to reduce the complexity and the computation time is to assume independence between vehicles. However this