The Research Of Multi-video Source Real Time Stitching

During the real-time video stitching, at one hand, there are many invariants in the video,such as scene or illumination. But most feature detectors just part of invariant about theseinvariants, and this will lead to the bad stitching directly. At other hand, high complexityalgorithm can’t use to real-time stitching since most image process contains mass ofcomputing. In order to achieve real-time stitching, not only the stitching algorithm must haverobust about these invariants, but also with the low complexity.Targeting the multi-video real-time stitching, this paper improved the originalalgorithms which are not suitable for real-time stitching. The main works of this paper asfollows:(1) Since SURF just has part invariant for illumination, and the illumination in the videocould be changed at any moment, we added the illumination invariant to the SURF, so theSURF can fully invariant for illumination and suitable for the feature detection of thereal-time video.(2) In order to reduce the computing of panorama stitching, this paper use video frameinterpolation strategy to achieve real-time video stitching. Since the path-based interpolationhave best interpolation output but with high complexity, this paper combined the path-basedinterpolation with optical flow to reduce the size of the path set, and it reduced thecomplexity of the interpolation algorithm.(3) In this paper, we proposed a method about the multi-video real-time stitching anddivide this process into panorama stitching and video interpolation, and using the CUDAmodel to accelerate the multi-video real-time stitching processes for the GPGPU have theproperty of Multi-Core-Multi-Thread. Through the GPGPU parallel computing, we canoutput the panorama video for real-time. In addition, according to the research results of thispaper, a real-time binocular stitching system was achieved which achieved a real-time videostitching.