Abstract:
This thesis proposed methods for motorcycle and motorcycle license plate detection from real-life traffic images. Motorcycles were separated from other vehicles and backgrounds using motorcycle shape analysis and neural networks. If motorcycles were detected, they would be sent to the license plate detection step. Motorcycle license plates were detected using multi-part gray level mean analysis that consisted of two filters: a rough filter and a thorough filter both of which were used to approximately locate and confirm license plates, respectively. These methods were tested on both motorcycles with license plates and motorcycles without license plates. In addition, the algorithm was enhanced to reduce false detection of bicycle as motorcycle. The experimental results gave accuracy of 92.90 % for motorcycle detection, 91.49 % for motorcycle license plates, 83.00 % for motorcycles without license plates. The algorithm would not specify bicycles as motorcycles with an average accuracy of 83.64%.