Minimal Injury Risk Motion Planning Using Active Mitigation and Sampling Model Predictive Control

Abstract

Collision mitigation is an important element in motion planning. Although Advanced Driver-Assistance Systems (ADAS) have a rich number of functionalities, they lack interchangeability. There is still a gap on finding a way to evaluate the best decision globally. This paper presents a novel motion planning framework to generate emergency maneuvers in complex and risky scenarios using active mitigation. The classical Model Predictive Path Integral (MPPI) algorithm is improved to be used in a probabilistic dynamic cost map under limited perception range. A cost map with global probability of injury to all road users is used as a constraint to the problem in order to compute target selection based on the global minimum risk considering all road users. Real experiments introduce the use of augmented sensor data by merging simulation and real sensor data to safely produce collision and mitigation experiments. Results show that the proposed algorithm can perform correctly in real time on board of the vehicle, by finding collision-free trajectories in complex scenarios and compute viable target selection that minimizes global injury risk when collision is inevitable.

From Probabilistic Occupancy Grids to versatile Collision Avoidance using Predictive Collision Detection

“…As a pedestrian is about to cross the road ahead, the probabilistic risk assessment provides the clues for the planner to identify the risk of intercepting the trajectory of the pedestrian and to decide whether to safely cross before or to yield smoothly. This behavior of our avoidance system qualitatively compares with what has been presented in [13], [17] and [18]. pedestrian in predicted occupancy dangerous trajectories that intercept the pedestrian safe trajectories that cross before safe trajectories that deviate from path safe trajectories that yield illustration of sampled trajectories by their 3s future rear axle position (■: collision, ■: safe) Fig.…”

“…Unfortunately these methods do not consider uncertainties in the dynamic part of the environment. Only [17] considers probability of existence of dynamic obstacles. This work uses Bayesian dynamic occupancy grid with a MPPI planner to mitigate the collision damage in road scenarios.…”

“…Among the state-of-art collision avoidance systems that integrate both perception and navigation under uncertainty on dynamic occupancy grids, we have identified on the one hand [13], [18] that handle navigation in human-populated environments, [14], [15] and [16] that operate on structured multiple-lanes road scenarios and [17] that considers collision mitigation in road scenarios. These solution exploit prior knowledge of structure of the environment (some assume a multiple-lanes road structure) and of the obstacles (some assume that all obstacles are humans others that all obstacles are cars).…”