On time-memory trade-off for collision detection

Abstract

Abstract-Future collision avoidance systems, which are capable of fully controlling the vehicle, have to make critical decisions in a very short time. To do this, they need to check constantly if their own vehicle's occupancy collides with the other traffic participants' occupancy. Those collision checks consume a substantial amount of time and consequently, the collision avoidance systems could fail to intervene in complex scenarios. We propose a new approach to reduce the computation time for collision checks significantly. Instead of using geometric methods, we store finitely many possible collision scenarios between two objects in a table and thus collision checks become a matter of lookup table queries. To ensure that the finite number of configurations cover all possible scenarios, we use a novel abstraction technique which guarantees that every collision will be detected. The approach works for arbitrarily many traffic participants by applying the approach pairwise (own vehicle and other object) to each traffic participant. Randomly generated scenarios show that the new approach can be several times faster than geometric intersection techniques thanks to the trade-off between memory consumption and computation time.

Ideal Reference Point in Planning and Control for Automated Car-Like Vehicles

“…In general, two different approaches of collision checks exist depending on the representation of the environment. Either a discrete grid map like in [20] or a geometric representation of the environment like in [21] can be used. In the following we will focus on the low-level grid-based environmental model instead of a high-level geometric representation, which requires detection and tracking of objects.…”

Anytime Safety Verification of Autonomous Vehicles

“…Our proposed anytime procedure has been integrated into the open-source MATLAB R tool SPOT [21], which represents occupancies by polygons and implements Algorithm 1. Since collision detection involving polygons is relatively slow, we plan to speed it up using bounding volume hierarchies [23], [24] and pre-computed collision checks [25]. To generate a long-term trajectory, as described in Section III-B, we use the sampling-based approach in [26].…”