VIDEO FEATURE DETECTION AND MATCHING FOR STRUCTURE FROM MOTION SYSTEM
MetadataShow full item record
With the improvements in sensor technologies and image processing algorithms, computer vision has become a major tool for robots to recognize and gauge their surroundings. For instance, the Kinect sensor can be used as an excellent depth camera for indoor navigation. However, there exist situations that need recognition and spatial interpretation of the environment using limited hardware resources. The Kinect is not suitable for outdoor use, while LIDAR is too large and expensive to be installed on an autonomous miniature surveillance drone. Therefore, the use of a single camera is the only feasible option for many embedded applications. To perform SfM (structure from motion) by using single camera is challenging due to the complexity of 3D mapping. Feature detection and mapping is the very fist step to perform SfM. To be more specific, matched feature points are used as anchors cross images or frames. Without such matched feature points, most SfM method will not be able to generate reliable results; moreover, instead of using frames from videos as inputs, most feature detectors and matching strategies are designed for SfM applications using images as inputs. Therefore, this thesis will discuss how to detect feature points from video and match them effectively. Image projection and SfM fundamentals will be introduced in this thesis as well.