Visual Analysis in Traffic & Re-identification
Andreas Møgelmose · VBN Forskningsportal (Aalborg Universitet) · 2015
Automated analysis of traffic situations - be it cars, signs, or pedestrians - is becoming increasingly relevant and feasible with the advent of powerful sensors, computers, and algorithms. This PhD thesis tackles three themes within this realm: Traffic sign detection, pedestrian detection and analysis, and person re-identification. In traffic sign detection, the work comprises a thorough survey of the state of the art, assembly of the worlds largest public dataset with U.S. traffic signs, and work in machine learning based detection algorithms. It was shown that detection of U.S. traffic signs has traditionally lacked behind detection of European signs, which led to the effort of collecting the dataset and pushing the state of the art in detection performance for these signs by using the Aggregate Channel Features detector. Within pedestrian detection, a method combining Viola-Jones and HOG/SVM has been put forth, which gives the speed advantage of Viola-Jones and the detection performance of HOG/SVM. Work has also been done in tracking the gaze of drivers to determine which pedestrians a driver may have missed. Finally, pedestrian tracking has been performed in an attempt to predict their future behavior in order to avoid dangerous situations. Person re-identification has been attempted in a multi-modal fashion. Traditionally, re-identification has been performed using only RGB input from regular surveillance cameras, but we added depth and thermal information to the mix. Several iterations of a multi-modal system were tested, but the advantage of the additional information turned out to be limited.