Robust Video Background Estimation for Traffic Monitoring Based on the Singular Value Decomposition

Infrastructure-based surveillance can be used to supplement the safety concepts of autonomous driving. Possible sensors are static cameras for the detection of traffic and obstacles in the monitored scene. This thesis investigates how a robust image background model can be estimated, that enables the detection of foreground objects even if environmental influences deteriorate the image quality. Several innovations of a given algorithm based on the singluar value decomposition (SVD) are presented. Multiple color and regionbased pixel features increase the information stored in the background model. A novel fusion process of long-term and short-term models based on the position of the sun enables modeling moving background shadows. Additional calculation of the robust principal component analysis (RPCA) prevents the inclusion of foreground objects into the model. All proposed methods are modularly implemented in an Octave testbench and combined into a final robust algorithm. A newly developed environmental challenges benchmark experimentally proves increased robustness and helps to identify limitations. Compared to the given algorithm, foreground detection performance can be improved by 23%.