Detection of vehicles using fisheye cameras
MetadataShow full item record
This item's downloads: 574 (view details)
An image processing system capable of detecting vehicles within the vicinity of the ego vehicle for use in a surround view camera system (SVC) is presented in this thesis. Image processing based object detection can be challenging when working with fisheye cameras due the effect of changes in perspective of the object combined with image distortion. In the proposed system, several feature-based vehicle detection methods are used to augment an industry-standard AdaBoost classifier. These vehicle detection methods are deployed in image regions where they have been found to perform optimally, thus ensuring the increased field of view offered by the SVC system can be fully exploited for vehicle detection purposes. The Hough Circle Transform (HCT) is used to detect target vehicle wheels due to their circular appearance when the vehicle is situated laterally with respect to the ego vehicle. A novel approach entitled Wheel Arch Contour Detection (WACD) detects target vehicle wheels in regions to the front/rear of the ego vehicle where both AdaBoost and HCT based detection methods produce weaker detection rates. Ellipse Based Wheel Detection (EBWD) is also used in specific circumstances to ensure continued detection between the AdaBoost classification and the HCT based techniques as well as side views.\par Harris Corner/Optical Flow techniques are used as visual tracking techniques in the event of detection not occurring in a given frame. Kalman filter prediction provides additional support for visual tracking. Novel and robust testing methodologies have also been developed to ensure spatial accuracy across the entire field of view of the SVC system. Such methods quantify performance of the proposed detection methods as a function of spatial displacement of the ego vehicle. The proposed improvements are useful when developing detection algorithms involving automotive sensors. Thorough testing has been carried out using a large dataset of various driving environments and has proven the effectiveness of the proposed approaches across the entire SVC field of view.