| Summary: | 博士 === 元智大學 === 電機工程學系 === 102 === How to detect and analyze vehicles is one of important issues in security surveillance and intelligence transportation system application. Nowadays, there are many cameras and surveillance systems installed in the intersections of cities to monitor passing vehicles. The vehicle information provided from these systems is very useful for policemen in many applications such as security monitoring, counter-terrorism, and etc. However, current surveillance systems limit in recognizing only license plate numbers. In addition to license plates, other information such as vehicle types and colors are also important for policemen to search suspected vehicles. This paper will address various challenges in vehicle make and model recognition, and vehicle color classification.
This paper proposes a new symmetrical SURF descriptor to enrich the power of SURF to detect all possible symmetrical matching pairs for vehicle detection and analysis. Without using any motion features, each vehicle can be located extremely accurately and efficiently from the matching results even though only single image is handled. Two advantages can be gained from our proposed detection scheme, i.e., (1) no need of background modeling and subtraction and (2) extreme efficiency for real-time applications. After that, two challenging tasks are handled in this dissertation; vehicle make and model recognition (MMR), and vehicle color classification. Generally, there are two challenging tasks in MMR, i.e., the multiplicity and ambiguity problems. The multiplicity stems from the small amplitude modification of same type vehicle by the manufacturer. The ambiguity results from different type vehicles which often share similar shapes. To tackle these two problems, a grid division scheme is proposed to divide a vehicle into several grids. These different weak classifiers are trained individually, and then combining these weak classifiers builds a strong ensemble classifier. The ensemble classifier can accurately recognize each type vehicle. As to vehicle color classification, a novel color correction method is first proposed to compensate the colors of a vehicle to maintain its color constancy. Then, we adopted a novel tree-based classifier to classify vehicles into detailed categories with significant accuracy improvement by dividing the root classifier to multiple root ones with several part-classifiers. Experimental results prove the superiorities of our proposed method in vehicle MMR and color classification.
|
|---|
