Environmental perception

Environmental perception

Predictive sensor technology (e.g. radars) enables more adaptive activation of airbag systems briefly before a collision. The prerequisite is reliable and fast detection of the near vehicle environment in all traffic situations - even under adverse environmental conditions. The challenge is to maximize accuracy while minimizing sensor cycle times, considering weather effects on sensors (e.g. rain, fog) and in highly dynamic driving situations (e.g. skidding before a collision). For this purpose, new signal processing methods are researched that determine the crash-relevant parameters of an impending accident with another vehicle or a pedestrian.

Wheel detection

The increasing resolution of environmental sensors (e.g. radar sensors) over the last decade enables improvements in the quality and details of object recognition. Especially in the pre-crash phase in which the distance between sensors and potential collision objects is small, additional information can be collected and evaluated. Particularly useful are object features whose positions are fixed within the object, e.g. the positions of the vehicle wheels. These distinctive points enable a reliable object recognition and tracking.

Therefore, a method is developed in C-ISAFE to determine the position and velocity of rotating wheels based on the micro-Doppler effect. The additional information is fused with data from camera and lidar and provides a robust and improved description of the detected objects.

Contact: Prof. Dr.-Ing. Thomas Brandmeier, Dr. Dagmar Steinhauser

Robustness against environmental influences

To increase safety and comfort through driver assistance systems up to automated driving, the reliable use of environmental sensors is mandatory for any environmental condition. In C-ISAFE, the performance of sensor systems is tested under reproducible boundary conditions in a defined test environment by using weathering facilities. Depending on the environmental influence (rain, fog, light), different types of disturbances are identified and characterized for the different sensor types (camera, radar, lidar). Methods for the reduction of disturbances on sensor systems as well as models for describing noise parameters are developed in order to enable an optimal environment perception.

Contact: Prof. Dr.-Ing. Thomas Brandmeier, Dr. Dagmar Steinhauser

Robust object tracking

Almost every 6th accident involving a passenger car with injured occupants shows skidding in the pre-crash phase despite modern ESP systems (source: GIDAS). Causes for this include inappropriate driver intervention in safety-critical driving situations or a slippery road surface due to snow and ice. Skidding represents a special, non-linear vehicle motion, which strongly depends on external environmental conditions (e.g. friction coefficient of the road surface).

In order to be able to activate future safety systems also in these driving conditions by forward looking sensors, the near vehicle environment must be reliably and continuously detected. In addition, the activation of passive safety systems (e.g. airbags) represents a safety-critical decision and is classified the highest Automotive Safety Integration Level (ASIL). Therefore, to increase robustness and safety, independent plausibility methods of environment information from forward looking sensors are developed and researched with new methods of object detection and tracking in critical driving situations. These methods are implemented on prototype vehicles and tested under real environmental conditions in C-ISAFE.

Contact: Prof. Dr.-Ing. Thomas Brandmeier, Robert Naumann

Pedestrian detection

About half of all traffic fatalities worldwide occur among vulnerable road users (VRUs) such as pedestrians and cyclists. VRUs always denote the weaker collision partner in road traffic. Therefore, the sensor-based detection of VRUs is crucial to optimally support the driver and thus increase the safety of all road users. Especially with regard to autonomous driving, robust detection in any traffic situation is indispensable. This includes unclear, complex scenarios as well as in the event of bad weather conditions.

In C-ISAFE, high-resolution radar-based methods are investigated to enable robust and fast classification of VRUs. By evaluating the micro-Doppler effect (velocity distributions due to the extremities’ motions) as well as high-resolution distance information, essential insights into the VRU’s behavior can be perceived in addition to pure classification. Motion indications are investigated to detect changes in the motion sequence, e.g., change of direction, at an early stage. Subsequent methods are developed to utilize those indications for path predictions. This is especially important since the pedestrian’s behavior is crucial in determining whether a dangerous situation arises. Consequently, path predictions for all relevant objects is essential to be able to react early in a dangerous situation.

Contact: Prof. Dr.-Ing. Thomas Brandmeier, Patrick Held

Number plate recognition

The THI License Plate Dataset (TLPD) was originally created to support the development of an algorithm capable of anonymising license plates collected with a dashboard camera. Today, the data set contains more than 17,000 vehicle images and approximately 18,000 labelled number plates taken at various angles, distances, lighting and weather conditions.

For more information please follow the link: www.thi.de/go/license-plate-detection

Contact person: Prof. Dr. Alessandro Zimmer