Tuesday, 29 January 2019

Raising the IQ of sports production



FEED


Mobile Viewpoint's AI-driven production solution makes broadcasting possible for any sport anywhere.


The evergreen popularity of live sports is part symptom and part cause of the rapid adoption of video streaming services but while the demand is huge, only half of the opportunity is being seized. Accelerating output to keep up with content-hungry sports fans is the next big challenge and one that automated production systems can help deliver.
“In the same way we have seen costly satellite trucks replaced by backpack-sized live video transmission units, AI will deliver similar cost savings to live production and streaming,” says Michel Bais, managing director, Mobile Viewpoint, developer of a system for sending video over the internet using mobile networks.
Whether that’s by using AI to analyse audio signals, video images, people or objects to identify which cameras to switch to and control, removing the need for an expensive camera crew, or from algorithms that can generate replays and graphics and then live stream content, AI (augmented or assisted intelligence, however you describe it) looks set to revolutionise the industry.
Mobile Viewpoint is not alone in looking to extend its capabilities beyond the competitive cellular uplinks market. Canada’s Dejero, for example, has a mobile app to helps bridge the gap between traditional remote production and cloud production while TVU Networks has also identified automatic sports logging and production as a related field for its core bonded cellular units.
The Dutch developer, though, has arguably gone further than any in co-opting AI into the workflow. It’s worked with research institute TNO to retune an off-the-peg AI brain from Google’s TensorFlow to fit the patterns, behaviours and relevant data sets of live streamed football.
The package, branded IQ Sports Producer, comprises two parts and includes a 32MB dome camera which contains four lenses, the feed from which is stitched into a 180-degree panoramic field of view. De-warping and stitching technology provides a corrected image with straight line markings. This image can be cut out, panned, or zoomed into as directed by the AI-software which is trained to follow the action as a human director would. All this takes about 30 frames, or 1 second delay.
At its crudest the AI will track groups of players (being most likely where the action is) but more sophisticated applications can be taught to anticipate where a ball might be passed to.
The camera itself was originally built by China’s Hikvision for surveillance purposes but has proved ideal for this solution, not least because it is relatively cheap (around Euro 5000 per module) and has been designed to withstand all weather as well as vandalism. Plus, the camera’s 4 x 4K feeds are reliably in sync which, Bais says, is not always the case with similar cameras.
The second part is hardware containing three Nvidia GPU boards, the AI software and the mobile connectivity which for Mobile Viewpoint is a no-brainer. “IP links are our bread and butter,” says Bais.
The AI software further combines motion-tracking with positional and other biometric data gathered from sensors (RFID, accelerometers, GPS) worn on special vests by the players themselves.
The first (currently only) system with the 32MP camera in Europe is installed at AFC Ajax, the biggest club in the Netherlands and part of the UEFA Champions League. Ajax is using it to film and stream its matches to club web channel Ajax TV. The footage is also captured to gather data on player performance analysis during training.
The Ajax training academy can monitor why a player missed a goal, why they failed to make an assist, and help improve their performance. Image detection means the AI technology can recognise different players and follow them, or, detect a ball on a pitch and follow its movements.
“The potential for AI in this respect is huge, especially in the production of live sports content,” says Bais. “As algorithms develop, AI can detect faults (yellow or red cards) or injuries as it learns how to make productions more interesting and story-like. There are plenty of smaller sports that could use this technology to become content owners in their own right at a low cost, and then monetise it.
“Only 10 per cent of professional sports are premier or first leagues. The majority are second or third tier sports many of which are played at grounds or parks with limited internet connectivity, even in middle of a city.”
That’s where the bonded cellular links come into play. It’s 3G and 4G capable today, encoded in HEVC H.265, and primed for 5G which telcos worldwide including local giant KPN will rollout in 2020.
The Netherlands Pro women’s football league, for example, is keen to install the system at six venues throughout the country as a budget-friendly means of launching a new online matchday streaming service.
However, to completely automate live sports streaming at the top level without the need for a production crew and director is perhaps a bridge too far—for now. AI used for assisted production alongside humans is the immediate goal. One of the obstacles to fully automated streaming is that it simply takes time for an algorithm to learn the nuances of what is interesting or important for each sport. For example, an algorithm may think that capturing a fight breaking out during a football match is the same as capturing a punch being thrown during a boxing match—but for the viewer, these are two very different experiences—one that is normal, and one that isn’t.
“Getting the algorithms up to speed requires time, so there is still very much a role for humans,” says Bais. “It tends to look for people walking or running rather than people on the ground so it doesn’t necessarily zoom in on a player who went down following a tackle which is maybe what the viewer would like to see.”
Higher image resolutions help too, since the more pixels the AI has to play with it can, for example, more easily differentiate the ball from a piece of white paper on the pitch.
While most budget-conscious customers want a single camera solution, more cameras around the venue are needed to capture different angles and more data to achieve greater accuracy.
Planned enhancements to the software include AI advertising insertion, auto playback of replays following a goal, auto summary (highlights) generation and to train the system on other sports including motorsport and ice hockey.
It has even created a means of translating the game into 3D VR graphics within which a user can select the point of view of any player to analyse decision making during a game. Currently used by Ajax as a training tool, “our ambition is to bring this to the home too,” says Bais. “With KPN we are working to bring a layer of interactivity to live streamed games running on KPN’s standard set-top box.”


No comments:

Post a Comment