| Summary: | 碩士 === 國立臺灣科技大學 === 電機工程系 === 106 === Recent developments in the area of deep learning and GPU computing have
benefited researchers to obtain information rich features to tackle various computer vision and localization problems and have aided in the development of various smart assistance systems such as driver assistance systems, surveillance systems etc. to ensure safety. This thesis presents a single camera based Advanced Driver Assistance System (ADAS) for navigation in highway. The main objective of this project is to develop a robust lane detection and vehicle identification system. Random Sample Consensus (RANSAC) algorithm is used to detect the road lane boundaries. To detect vehicles in the scene, a deep learning based object detection algorithm called Single Shot MultiBox Detector (SSD) is used. This detector is based on a feed forward network which produces a set of bounding boxes for an object and scores for the presence of the object within those boxes. For this project, the network is trained with KITTI benchmark dataset. This algorithm can process up to 59FPS. With known parameters such as the height at which the camera is mounted,tilt angle of the camera and bounding box obtained using the deep learning vehicle detection system, the distance between vehicles detected in the scene and the camera can be computed. This is done using a series of trigonometric relationships between the image formed in the camera and the real world scene. A camera calibration method is also proposed to calculate few essential parameters for distance estimation.
|
|---|
