Introduction
Path prediction is a key pillar for systems like Advanced Driver Assistance Systems (ADAS) or Cooperative Intelligent Transport Systems (C-ITS) to recognize dangerous situations ahead. For motorcycles (PTWs), path prediction will help to reduce accidents. At the same time, to predict the future path of a PTW, challenges remain: The vehicle dynamics of PTWs differ fundamentally from cars and are more complex.
What has been investigated
This white paper document is intended to give an overview of possibilities with the current technical state of the art, related to different sensors and algorithms. It further highlights limitations regarding curvature calculation and path prediction on PTWs.
CMC decided to divide the topic of path prediction into three levels, and focus in this white paper on two of these: Level C and Level B.
* Level C path prediction covers straight line riding/driving; Based on information provided by the standard Cooperative Awareness Messages (CAMs). Level C is achieved by constant velocity, heading and position using well-known information from CAMs.
* Level B path prediction covers steady state (constant curvature) riding/driving. In addition to the information that Level C uses, Level B uses constant turn rate, additionally the radius of curvature calculated by the ego vehicle.
In both cases the path predication and collision risk can be calculated based on a ‘Ghost Vehicle’ method.
A consideration is made as to what technological advances would be needed to make acceptable path prediction.
These could be heading in the direction of level A path prediction: advanced and complex methods such as contextual behaviour predictions using machine learning and advanced sensing technology such as cameras and radars.
For this white paper, CMC has decided to focus on a typical ‘left turn’ scenario and use it as an example for path prediction, since this case often involves severe or even fatal consequences.
CMC decided to divide the topic of path prediction into three levels, and focus in this white paper on two of these: Level C and Level B.
* Level C path prediction covers straight line riding/driving; Based on information provided by the standard Cooperative Awareness Messages (CAMs). Level C is achieved by constant velocity, heading and position using well-known information from CAMs.
* Level B path prediction covers steady state (constant curvature) riding/driving. In addition to the information that Level C uses, Level B uses constant turn rate, additionally the radius of curvature calculated by the ego vehicle.
In both cases the path predication and collision risk can be calculated based on a ‘Ghost Vehicle’ method.
A consideration is made as to what technological advances would be needed to make acceptable path prediction.
These could be heading in the direction of level A path prediction: advanced and complex methods such as contextual behaviour predictions using machine learning and advanced sensing technology such as cameras and radars.
For this white paper, CMC has decided to focus on a typical ‘left turn’ scenario and use it as an example for path prediction, since this case often involves severe or even fatal consequences.
Important outcomes
CMC examined how far ahead (how many seconds ahead) the calculated predicted path is valid, within a threshold of 2 meters.
Various intersection passage scenarios were investigated with a BikeSim tool and five different algorithms were tested.
In most cases, there were no big differences between the algorithms. During an actual manoeuvre, a path predication of only up to about 1-2 seconds ahead was the limit of these systems. On straights, the path predication can be valid for more seconds, but as soon as there is a small degree of roll motion (which is natural for PTWs, even on straights) this is already reduced to 2-3 seconds.
The main factor behind the performance limit is that it depends fully on the instantaneous curvature. In order to improve that, consideration could be given to items like rider intention, map information, and next algorithm improvements; in the future also machine learning / artificial neural networks.
For the moment, however, this means that an oncoming vehicle driver cannot be informed early enough whether the PTW is going straight or turning and path prediction for PTWs is not sufficiently solved.
Various intersection passage scenarios were investigated with a BikeSim tool and five different algorithms were tested.
In most cases, there were no big differences between the algorithms. During an actual manoeuvre, a path predication of only up to about 1-2 seconds ahead was the limit of these systems. On straights, the path predication can be valid for more seconds, but as soon as there is a small degree of roll motion (which is natural for PTWs, even on straights) this is already reduced to 2-3 seconds.
The main factor behind the performance limit is that it depends fully on the instantaneous curvature. In order to improve that, consideration could be given to items like rider intention, map information, and next algorithm improvements; in the future also machine learning / artificial neural networks.
For the moment, however, this means that an oncoming vehicle driver cannot be informed early enough whether the PTW is going straight or turning and path prediction for PTWs is not sufficiently solved.
Documentation
Scroll through the complete document here below, or download it as PDF: