Real-time Person Tracking in High-resolution Panoramic Video for Automated Broadcast Production

Real-Time Person Detection and Tracking in Panoramic Video

“…The approach extends the tracking by detection approach we have proposed in [6] by addressing a number of robustness issues that are especially relevant in sports content (e.g., fast motion, motion blur, occlusions). The design of the approach is strongly driven by the requirement to process this highresolution data in real-time, thus we propose a number of performance optimizations.…”

“…The approach extends the tracking by detection approach we have proposed in [6] by addressing a number of robustness issues that are especially relevant in sports content (e.g., fast motion, motion blur, occlusions) and optimizes some steps in the process. We first briefly summarize our earlier work.…”

“…To connect person detections of subsequent frames, a KLT tracker implementation based on CUDA [3] is used. In [6], a further optimization taking advantage of NVIDIA's Fermi 3 architecture is described, resulting in real-time tracking capability of up to 10K feature points in two FullHD streams on a single GPU. Figure 1.…”

“…Figure 1. Overview of the tracking by detection algorithm described in [6] Because of the tile-based processing, the approach can be parallelized in order to achieve overall real-time performance. This can either be done by processing tiles on separate machines or one a single machine, with a multi-core CPU and multiple CUDA enabled GPUs.…”

Tiling of panorama video for interactive virtual cameras: Overheads and potential bandwidth requirement reduction

“…For example, tiling is discussed in [6], [7], [8], [9], [10]. Furthermore, [11], [12], [13], [14], [15] extensively address the problem of automatically generating personalized content, and [16] discusses plain cropping. 1 This size depends on the lighting and weather conditions.…”