Optimized Distance Measurement with 3D-CMOS Image Sensor and Real Time Processing of the 3D Data for Applications in Automotive and Safety Engineering
Bernhard König · DuEPublico (University of Duisburg-Essen) · 2008
This thesis describes and characterizes an advanced range camera for the distance range from 2 m to 25 m and novel real-time 3D image processing algorithms for object detection, tracking and classification on the basis of the three-dimensional features of the camera output data. The technology is based on a 64x8 pixel array CMOS image sensor which is capable of capturing three-dimensional images. This is accomplished by executing indirect time of flight measurement of NIR laser pulses emitted by the camera and reflected by the objects in the field of view of the camera. An analytic description of the measurement signals and a derivation of the distance measuring algorithms are conducted in this thesis as well as a comparative examination of the distance measuring algorithms by calculation, simulation and experiments; in doing so, the MDSI3 algorithm showed the best results over the whole measurement range and is thus chosen as standard method of the distance measuring system. A camera prototype was developed with a measurement accuracy in the centimeter range at an image repetition rate up to 100 Hz; a detailed evaluation of the components and of the over-all system is presented. Main aspects are the characterization of the time critical measurement signals, of the system noise, and of the distance measuring capabilities. Furthermore this thesis introduces novel real-time image processing of the output data stream of the camera aiming at the detection of objects being located in the observed area and the derivation of reliable position, speed and acceleration estimates. The used segmentation algorithm utilizes all three spatial dimensions of the position information as well as the intensity values and thus yields significant improvement compared to segmentation in conventional 2D images. Position, velocity, and acceleration values of the segmented objects are estimated by means of Kalman filtering in 3D space. The filter is dynamically adapted to the measurement properties of the according object to take care of changes of the data properties. The good performance of the image processing algorithms is presented by means of example scenes.