Development of video stitching algorithms

Abstract

Video stitching is a process to generate a wide field of view video from multiple over-lapping videos captured from ordinary cameras. Currently, video stitching has become a popular research topic in computer vision. Video stitching is being used in surveillance, virtual reality, augmented reality, virtual tour, games, and movies. Many algorithms for video stitching rely on locating feature points in each image, calculating feature descriptions from a region of pixels surrounding each feature point. This research work is also based on this approach for providing effective presentation of videos. In this thesis, the challenges faced while stitching the videos from multiple cameras are addressed. While stitching the videos from multiple cameras misalignment, artifacts and ghosting effects occurring due to parallax error need to be addressed. Apart from this challenge, Spatial-temporal consistency also needs to be tackled. Spatial-temporal smoothness is affected by the motion of camera as well as by the motion of objects between the frames. Videos, in general, captured by moving cameras appear shaky, which will affect temporal smoothness. The proposed work focuses on developing video stitching algorithms with Interacting Multiple Model feature tracking technique, spatial-temporal region-based method, Stationary wavelet transform based feature detection method and Multi view spatial-temporal feature method. In the Interacting Multiple Model Feature Tracking technique, the feature-based video stitching algorithms mostly rely on tracking of feature points. In this work, the spatial-temporal consistency is maintained by designing a model that accurately track the feature points and correctly newline