Automated Analysis of On-Board Network Data
Today, 90% of the innovations in vehicles can be contributed to the fields of electronics and mechatronics. This trend, which will continue in a strengthened way, comes along with an increasing software part in the electronic control units (ECU). Along with a permanently increasing degree of networking in the vehicle this leads to a more complex development, validation, and production. So, most of the vehicle functions which the driver experiences are no longer controlled by a single ECU, but provided through the integration of subsystems and, thus, through the networked interaction of the individual sub-functions at the right time. Securing the cooperation of these functions at the right time is already one of the key tasks of the electrics and electronics validation today. To cope with this task, the communication between the ECUs must be recorded. The rapidly increasing number of vehicle data bus systems (e.g., CAN, LIN, FlexRay, MOST) leads to a huge amount of recorded data. For example, the amount of data recorded per hour in a new BMW 7series is seven times greater than the volume of data at the previous 7series.
It is not possible to analyze and evaluate these huge amounts of network data in a completely manual way. Consequently, our framework for automated on-board network analysis provides methods to check these large amounts of data automatically against the reference and the erratic vehicle behavior which is given in the form of rules. Therefore, vehicle defects, suspicious facts, and irregularities in the recorded data can automatically be shown. The required rules are created from recorded data by means of techniques from the fields of computational intelligence and data mining. As an alternative, they can also be generated from existing specifications or created manually. The overall process includes the preparation of the recorded data, the semi-automatic analysis to support the work of the analysis specialists, the fully automated verification of records by a so-called analysis engine, and the corresponding result documentation to provide a feedback to the application engineer.
Together with the BMW Group, we develop techniques to identify very rare errors in records of vehicle network data.
Collaborative Situation-Awareness of Vehicles
Driver assistance systems are a key issue in the visionary field of accident-free driving. Direct communication of vehicles allows for a collaborative situation-awareness. Vehicles can be enabled to predict critical driving conditions and can inform their drivers in a timely manner. However, such systems typically have to deal with incomplete and unconfident information. In addition, spatial and temporal effects have to be taken into account.
Together with the BMW Group, we developed a novel approach for collaborative situation-awareness of vehicles based on Bayesian networks. This approach allows for an analysis of remote locations on the road ahead using both, on-board sensor data and observations from other vehicles.
Object Classification for Automotive Safety Applications
Driven by the growing demands for road safety systems, more and more development activities in the area of intelligent vehicles are focusing on active and/or passive safety applications to build a virtual safety belt around the host vehicle in order to warn or respectively protect the passengers as well as vulnerable road users (like pedestrians, cyclists, etc.) in case of dangerous situations or accidents. The key challenge for these driver assistance and safety systems is the accurate perception of the vehicle’s surrounding with a high reliability and measurement precision.
In the majority of cases this task is addressed with data fusion of several car mounted sensor devices. On the basis of their data the perception system has to detect and track respective objects in the vehicle’s environment. In addition to the object tracking, an object classification is essential for most automotive safety applications: The protection of vulnerable road users by means of autonomous braking needs a robust identification of road users, for instance.
Object classification in automotive applications is a research focus of FORWISS (Passau). We contributed techniques based on support vector machines to classify low level sensor data from radar sensors and a laser scanner with the purpose to identify various kinds of road users.