Detecting pedestrians intending to enter a crosswalk using a HMM tracker and a novel predictor

Sammanfattning: There is a demand for a more fluent and more efficient traffic. This can be achieved through more intelligent traffic light control. This thesis presents theory and an application in which people are tracked and their intentions to cross a crosswalk is predicted with a novel prediction algorithm based on Markov theory. The background segmentation and tracking algorithms was based on already known cross-correlation and HMM-methods. Based on the relatively small amount of training data the result for the novel predictor detecting persons "entering the crosswalk" for two different setups, a straight and a four way crossing, is 70% and 55% true positives with 5% and 2% false positives. For detection of someone that is "not entering the crosswalk" when there is a person in the area is 90% and 85% true positives with 15% and 25% false positive. The results achieved are good enough as a proof of concept that the theories are worth investigating further for these kind of applications. However, a lot of work would still be required before this is robust enough to be in a real traffic application.