Abstract
Keywords - 3D-Camera, Object Detection, Machine Learning, PMD
BMW already has several sensors in series production to support safety systems. Examples are the night vision FIR camera and the radar for the Active Cruise Control (ACC) system. These sensors model the surroundings of the car and the data is processed by the assistance systems to enable the driver to experience more "Sheer Driving Pleasure" but also be safer in critical or uncomfortable situations. Another sensor in research is a PMD (photonic mixer device) camera which is based on the time-of-flight principle to measure distances and returns within one pixel an intensity and a distance value. So there is a unique and new technique for capturing 3D data from the environment to segment and classify objects such as pedestrians, cars, trucks or road side objects.
The segmentation described in this paper is done through a region growing algorithm which groups pixels having the same distance from the sensor.
Classification can be done with several different methods. One way is to define rules beforehand and apply them to the given data, e.g. a car has a certain width or height and intensity in an image and is by these features then classified as a car in the 3D data. However, often there are many features that have to be taken into account and can influence the classification differently. For this situation machine learning algorithms such as Support Vector Machines, Decision Trees and K-Nearest-Neighbor can be used. For these algorithms, supervised learning is applied where several features are extracted from a training set and mapped to the different object classes such as pedestrians, cars, trucks and other objects. After this training step the collected information and knowledge can be used to classify given data. This paper describes the use of machine learning algorithms on data from a time-of-flight camera for the classification of the object segments given by the implemented segmentation algorithm. There will be a detailed explanation of the used algorithms and features and also an evaluation of the performance of the classification.
The classified segments can then be used for various safety applications.