WP6: Anomaly and situation detection for traffic safety

Dr. Anders Dahlbom, WP leader 
Dr. Maria Riveiro, Dr. Alexander Karlsson, Prof. Sten F Andler
M.Sc. Emma Johansson, M.Sc. Marcus Elmer, Dr. Malte Ahrhold, M.Sc. Hui Zhong

A goal in the area of traffic safety applications is to reduce the number of accidents that occur. Active safety systems are concerned with inferring hazardous driving behavior or situations before they occur so that they can be prevented. This is an impact analysis problem that in principal can be put in terms of classification, where time series data is used together with some classification function to determine if the present behavior or situation is hazardous. Classification can be carried out by using models of normal driving behavior and models of hazardous driving behavior. The problem is twofold: (1) construction of models and (2) utilization of models for classification. The former part can be addressed using both knowledge- and data-driven approaches. Although both types of approaches in theory can be used for detection, they have different properties that may affect their suitability. Although the detection of hazardous situations ultimately is carried out in an online system, to be prevented, it is also of great interest to be able to identify hazardous driving situations in historical data, in order to e.g. improve existing functions, for introducing new functions, or for analyzing accidents.
In a live system, these types of problems are addressed using advanced driver assistance (ADAS) functions, which are also used for e.g. increased comport. In an abstract view, ADAS functions can be structured into a perception of the vehicle’s environment, a situation assessment and finally a decision on warnings. Initially, it is based on a set of measurements that are subject to uncertainties, e.g. measurement errors, classification errors or the risk of false or missed detections. An important question concerns how to properly represent the uncertainties in the different processing steps and how to propagate them towards the decision mechanisms.
The main focus of this WP is to investigate anomaly and situation detection for traffic safety applications. A pre study with respect to uncertainty management for ADAS functions will however also be carried out.
________________________________________

Research Question

The following questions are addressed in this WP:

  • Which variables/states are crucial for detecting safety critical traffic situations?
  • Which methods are efficient for detecting safety critical traffic situations, while minimizing the number of false alarms?
  • Can interactive visualization be used to improve performance of anomaly detection?
  • In what way should knowledge driven and data driven methods be combined in order to be used for efficient detection of safety critical traffic situations?
  • How can uncertainty management methods be included in ADAS data processing?

Relevance to UMIF

This work package is an industry oriented work package that has close connections with all three of the more general work packages (1, 2 and 3). The problems studied in the work package touches upon uncertainty management, visualization, and impact analysis and anomaly detection, from a traffic safety perspective. The work package is thus clearly interesting for the intersection of these three areas that UMIF constitutes.

Collaboration with Industry

Since this is an industry oriented work package, there is a close collaboration with industry, in this case Volvo Technology. Two of the work package members from the University try to regularly work on site at Volvo Technology and address the questions in collaboration with industry participants. 

Collaboration with Academia

This work package collaborates with work packages one, two and three. Additionally, results should also be delivered to work package eight.

Approach

Due to the two tracks in this work package, two approaches are used. In the first track, data- and knowledge driven modeling is investigated using historical data from real driving situations. The second track investigates the impacts of different uncertainty management methods in a specific ADAS function.

WP Results & Status

This work package has been delayed and does therefore only have preliminary results. More time is planned to be allocated in the end of 2012 and in early 2013.
Two publications are planned for next year.

Related Work

The work carried out in this work package is strongly related to work packages 1, 2 and 3.


Uppdaterad: 2013-05-22
Sidansvarig: Marcus Brohede