Abstract
Latest accident analyses and statistics, as published in [1], clearly show a steady decrease of the number of injured and killed persons in night-time traffic accidents over the last decades. This positive trend is partly due to the continuously improving automotive technology with respect to passive and active safety features for both occupants and pedestrians. In particular, intensive accident analyses carried out by expert teams, as described in [2, 3], provide distinctive information which driving assistance functionality is to be invented or optimized further to assist the driver at night time. Apart from speed-related accidents driving situations in which pedestrians, cyclists or game animals are directly involved, have been seen as very critical. Especially for collisions with pedestrians, which are likely to happen in rural areas, a very high situational mortality can be observed. In [4] a novel intelligent lighting-based assistance system has been presented and named “Marking Light”, which has been inspired by these negative accident reports. In this context, objects which have the potential for a collision are detected by optical or thermal sensor systems, tracked over a certain period of time and finally classified. If classification has been positive a strategic analysis will be carried out in order to decide whether a collision within the so-called Time-to-Collision (TTC) is likely or not. If this gained value is above a specified threshold the driver’s attention can be positively influenced by using a dedicated lighting pattern (shape, lighting frequency) to avoid the accidental collision.
In this paper, an overall system for evaluating purposes will be described in detail, which will be used as a platform to carry out field tests in the future. On the sensor side a far infrared camera system, an inertial measurement unit as well as interfaces for CAN communication have been integrated. On the actor side we utilize a prototype headlight system, which integrates a dedicated lighting system based on high-power LED emitters for light-based marking purposes besides a bi-xenon module.
The main part of this paper will focus on the detection and classification of objects (e.g. pedestrians, cyclists, game animals, etc.) to be marked. For this purpose, details on the applied detection and classification framework will be provided, which integrates multiple trained instances of the so-called “Support Vector Machine (SVM)” algorithm. By means of a downstream meta-classification stage the most likely type of object class, which is essential for a sophisticated marking strategy (cf. HMI design) can be determined. A tracking algorithm based on a mean shift approach finally minimizes the false positive rate of the overall system.
In the outlook the next steps of the research project will be described such as an analysis of the aiming accuracy, or the necessary reduction of dazzling the marked objects in a moving coordinate system.
Keywords: Advanced driving assistance systems (ADAS), automotive lighting systems, marking light, object detection and classification, support vector machines