This paper presents an automatic traffic surveillance system to estimate important traffic parameters from video sequences using only one camera. Different from traditional methods which classify vehicles into only cars and non-cars, the proposed method has a good capability to categorize cars into more specific classes with a new "linearity" feature. In addition, in order to reduce occlusions of vehicles, an automatic scheme of detecting lane dividing lines is proposed. With the found lane dividing lines, not only occlusions of vehicles can be reduced but also a normalization scheme can be developed for tackling the problems of feature size variations. Once all vehicle features are extracted, an optimal classifier is then designed to robustly categorize vehicles into different classes even though shadows, occlusions, and other noise exist. The designed classifier can collect different evidences from the database and the verified vehicle itself to make better decisions and thus much enhance the robustness and accuracy of classification. Experimental results show that the proposed method is much robust and powerful than other traditional methods.
|Number of pages||8|
|Journal||Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)|
|State||Published - 1 Dec 2003|