Stochastic MPC with Multi-modal Predictions for Traffic Intersections

Non-Conservative Trajectory Planning for Automated Vehicles by Estimating Intentions of Dynamic Obstacles

“…The behavior of the EV controlled by the proposed algorithm is simulated in CARLA, to discuss the properties and benefits of the novel combination of the IMM and SMPC. We compare the performance with two similar approaches presented in previous works: [6], in which the probability of each candidate trajectories of the DOs is estimated but only the most likely is accounted for in the motion planning of the EV, and [31], in which multiple candidate trajectories are considered and regarded as equally likely. Urban environments with significant uncertainty concerning the future trajectories of dynamic obstacles are considered.…”

“…This approach allows to effectively consider multiple future behaviors, but relying on a single collision avoidance constraint that accounts for all candidate future motions at once requires significantly overapproximating the forbidden areas for the EV, and thus induces unnecessary conservatism. Differently, the recent work [31] considers several candidate maneuvers of DOs separately, avoiding the just-mentioned over-approximation. Still, all obtained predictions are regarded as equally likely, irrespective of the estimated probabilities.…”

“…In simulation 4, all candidate trajectories of the cyclist are taken into account with equally large safety areas, see Figure 8a. This approach is inspired by [31]. Initially, the EV behaves similarly to when controlled with the IMM+SMPC algorithm, that is, it maintains a distance and avoids overtaking at first.…”

“…We compare our novel IMM+SMPC algorithm with the two comparison methods from [6], [31] considering two metrics. On one hand, we evaluate the average computation time for each iteration of the IMM+SMPC method for the discussed simulations; specifically, in each iteration, we include the update step of the IMM, the multi-modal trajectory prediction algorithm and the generation of safety collision avoidance constraints, and the solution of the optimal control problem.…”

“…As a matter of fact, in this case at last the cyclist is predicted to occupy areas outside of the road boundaries, therefore the collision avoidance safety constraints are always satisfied in practice and the optimal control problem is solved quickly. Compared to our proposed IMM+SMPC method with the initial 3 candidate trajectories, the comparison method similar to [31], in which future candidate trajectories are regarded as equally likely, results in a slightly shorter average computation time, but in a 60% increase in the cost. Indeed, without prioritization of the candidate trajectories, the EV never excludes the possible lane invasion, even when the cyclist clearly shows a different intention.…”