| Summary: | 博士 === 國立成功大學 === 工程科學系碩博士班 === 100 === This dissertation addresses five techniques and applications for vehicular digital images, namely the vehicular digital video recorder system, the road sign detection and recognition, the driver fatigue detection and monitoring, the top-view transformation model for image coordinate transformation, and the forward collision warning system.
The proposed vehicular digital video recorder system is active during vehicle movement and can securely store the data. The system supports an online real-time navigator and an offline video data viewer. The data viewer subsystem can be used to analyze the recorded data and obtain the status of the vehicle after a traffic accident.
Adaptive image pre-processing models are proposed for a road sign recognition system that uses two fuzzy inference schemes. The first scheme is used to check illumination changes and rich red color of a frame image by the checking areas. The other scheme is used to check vehicle speed and the angle of the steering wheel to determine the size and position of the detection area. The Adaboost classifier is employed to detect road sign candidates from an image and the support vector machine technique is employed to recognize the content of the candidate road signs. Mandatory and warning traffic signs are the targets. The proposed system, which overcomes the problems of low illumination and the rich red color around road signs, has a high detection rate and efficiency.
Driver fatigue is one of the main reasons for traffic accidents. A driver fatigue detection system is proposed. The preprocessing stage includes face detection and eye position extraction. In the second stage, eye tracking is performed for a sequence of frames and the eye state is determined using a particle filtering scheme. The final stage executes the fatigue detection and monitoring, and issues warnings.
The proposed top-view transformation model for image coordinate transformation involves transforming a perspective projection image into its corresponding bird’s eye view. A fitting parameters search algorithm estimates the parameters that are used to transform the coordinates from the source image. The interior and exterior orientation parameters of the camera are not required for this approach. The designed car parking assistance system can be installed at the rear end of the car, providing the driver with a clearer image of the area behind the car. Experimental results show that the proposed approaches provide a clearer and more accurate bird’s eye view.
A forward collision warning system based on the vehicular digital video recorder system and a vision sensor is proposed. The forward collision warning system integrates edge detection, vehicle bounding box detection, and feature point matching techniques. An initial detection region of interest is given in a frame along with the corresponding vehicle speed. Feature points are detected and matched in two adjacent frames. The proposed vehicle bounding box detection approach filters more noise than existing methods and has high detection accuracy.
|
|---|
