Maneuver recognition in highway traffic

 
A highway, typically consisting of several traffic lanes, is characterized by complex traffic scenes involving many vehicles. To reduce the risk of accidents, a driver must interpret possible hazards of a situation accurately. This includes correctly recognizing intended maneuvers of all surrounding vehicles. If the driver is distracted or overloaded during the interpretation of a traffic scene, this can lead to accidents and to congestions causing CO2 pollution of the environment, besides macro-economical effects.

Identification and interpretation of traffic maneuvers will become key elements of modern driver assistance systems. Considerable effort has been put into early recognition of lane change maneuvers and a number of challenges have been identified. There are two main reasons why potentially dangerous maneuvers of neighboring vehicles represent a challenge for the early proactive recognition:
  • the situations develop quickly over time, and an automatic system will therefore require information captured in the order of milliseconds
  • the situations can only be reliably recognized when considering the joint behavior of several sensor measurements simultaneously, often featuring several vehicles.
The dataset, which is made available from the start of the project, includes both information on the motion state of the current and neighboring vehicles (e.g., position, speed, acceleration, orientation within the lane, trajectory, free space for a maneuver, see Figure 1), as well as information from the environment like lane markings and road borders. The situation features used for maneuver recognition are structured along three main dimensions (as shown in Figure 1): lateral evidence, trajectory, and occupancy schedule grid, for more details see [1], [2].

The automotive data-sets used in AMIDST are extremely large. Consider a highway scenario involving a vehicle driving in a lane with three other vehicles driving in three different lanes in front of the considered vehicle. The information describing such scenarios typically consists of 252 observations acquired with fixed sampling rate (in the order of milliseconds). If a test drive from only one hour is to be analyzed for adaptation of the model parameters, this will result in several millions of database records. (Using today’s sampling rate of 40 milliseconds, this gives 22.680.000 records.) For safety reasons, a modern driver assistance system is tested on drives involving thousands of kilometers with the corresponding high number of driving hours, which makes an off-line parameter adaptation of the models unfeasible due to a huge size of the database. This calls for efficient algorithms and methods that must be scaled up to handle the extremely large volumes of data (compared to the equipment available for on-line processing).



The AMIDST impact on future maneuver recognition in highway traffic
AMIDST optimizes the algorithms to fit the requirements on memory, inference time and computational power of the computer in the car. This optimization should meet the reactivity requirement and ensure that the systems developed must be able to operate at the time scale of the automotive processor they are designed to support. To realize this reactivity requirement, alternative modeling techniques and model implementations will be developed and tested for comparison.

AMIDST develops dynamic models to boost early recognition of potentially dangerous maneuvers based on observed and computed data streams. The model, equipped with a suitable inference engine to accommodate real-time decision-making under uncertainty, is expected to improve the ability to detect evolving and potentially dangerous maneuvers before they become critical. The advantage of a probabilistic approach to recognition of potentially dangerous maneuvers lies in the capability to handle the high level of uncertainty in the measured data.

AMIDST will provide a valuable contribution to achieve the goal EU27 on reduction of the number of fatalities in year 2011 by 50% until year 2020.




Figure 1: A: Symmetric lane-coordinate-system for each vehicle; B,C,D: Situation features for a lane change maneuver, in the right coordinate system (B: Lateral Evidence, C: Trajectory and D: Occupancy Schedule Grid)




References:
[1] D. Kasper, G. Weidl, T. Dang, G. Breuel, A. Tamke, and W. Rosenstiel. Object-oriented Bayesian networks for detection of lane change maneuvers. In Proceedings of the Intelligent Vehicles Symposium (IV), 2011 IEEE, 2011.
 
[2] D. Kasper, G. Weidl, T. Dang, G. Breuel, A. Tamke, A. Wedel, and W. Rosenstiel. Object-oriented Bayesian networks for detection of lane change maneuvers. Intelligent Transportation Systems Magazine, IEEE, 4:19–31, 2012.
 
 


This website uses cookies. Cookies are only used for handling login and identifying registrered users, not for visitor tracking.