Vehicle object retargeting from dynamic traffic videos for real-time visualisation

摘要

One form of video visualisation is to transform traffic videos from a street view to an aerial view, which facilitates a summary overview of multiple traffic video streams. This paper presents an efficient and effective solution to mitigate the undesirable distortion of the re-targeted vehicle objects in traffic video visualisation. This is achieved by a series of automated algorithmic steps, including vehicle segmentation, vehicle roof detection, and non-uniform image deformation by applying a second homography. This technique has been integrated into a video visualisation system that creates an aerial view of re-targeted video streams on top of a conventional aerial view. The results have shown that the technique offers the system a significant improvement in visual quality without undermining the requirement for real-time video visualisation.