System and method for calibrating moving camera capturing broadcast video
A system and method of calibrating moving cameras capturing a sporting event is disclosed herein. A computing system retrieves a broadcast video feed for a sporting event. The broadcast video feed includes a plurality of video frames. The computing system labels, via a neural network, components of a playing surface captured in each video frame. The computing system matches a subset of labeled video frames to a set of templates with various camera perspectives. The computing system fits a playing surface model to the set of labeled video frames that were matched to the set of templates. The computing system identifies camera motion in each video frame using an optical flow model. The computing system generates a homography matrix for each video frame based on the fitted playing surface model and camera motion. The computing system calibrates each camera based on the homography matrix generated for each video frame.
1. A method of calibrating moving cameras capturing a sporting event, comprising:
identifying, by a computing system, a live broadcast video feed for a sporting event, the live broadcast video feed comprising a plurality of video frames captured by a plurality of cameras remotely located in a sporting venue;
labeling, by the computing system via a neural network, components of a playing surface captured in each video frame to generate a set of labeled video frames;
matching, by the computing system, the set of labeled video frames to a set of templates, wherein each template of the set of templates represents a different camera perspective of the playing surface;
removing, by the computing system, at least one labeled video frame from the set of labeled video frames to generate a subset of labeled video frames, wherein the at least one labeled video frame does not match any template in the set of templates;
identifying, by the computing system, camera motion in each video frame using an optical flow model;
generating, by the computing system, a homography matrix for each video frame based on the camera motion; and
remotely calibrating, by the computing system over a network, each camera of the plurality of cameras based on the homography matrix generated for each video frame.
2. The method of claim 1 , wherein identifying, by the computing system, the camera motion in each video frame using the optical flow model, comprises:
generating, for each video frame of the plurality of video frames, body pose information for each player in each video frame; and
removing each player from each video frame based on the body pose information.
3. The method of claim 2 , wherein removing each player from each video frame based on the body pose information, comprises:
identifying one or more pixels corresponding to a player in a video frame based on the body pose information; and
removing the identified one or more pixels from the video frame.
4. The method of claim 1 , wherein retrieving, by the computing system, the live broadcast video feed for the sporting event comprises:
identifying, from the plurality of video frames, a set of trackable video frames, wherein each trackable video frame in the set of trackable video frames captures a same perspective of the playing surface.
5. The method of claim 1 , wherein generating, by the computing system, the homography matrix for each video frame based on the camera motion comprises:
training a further neural network to generate the homography matrix for each video frame based on the subset of labeled video frames and flow fields associated with the camera motion between successive video frames.
6. The method of claim 1 , wherein calibrating, by the computing system, each camera based on the homography matrix generated for each video frame comprises:
projecting player coordinates in each video frame to real-world coordinates based on the homography matrix of each video frame.
7. A system for calibrating moving cameras capturing a sporting event, comprising:
a processor; and
a memory having programming instructions stored thereon, which, when executed by the processor, performs one or more operations, comprising:
identifying a live broadcast video feed for a sporting event, the live broadcast video feed comprising a plurality of video frames captured by a plurality of cameras remotely located in a sporting venue;
labeling, via a neural network, components of a playing surface captured in each video frame to generate a set of labeled video frames;
matching the set of labeled video frames to a set of templates, wherein each template of the set of templates represents a different camera perspective of the playing surface;
removing at least one labeled video frame from the set of labeled video frames to generate a subset of labeled video frames, wherein the at least one labeled video frame does not match any template in the set of templates;
identifying camera motion in each video frame using an optical flow model;
generating a homography matrix for each video frame based on the camera motion; and
remotely calibrating, over a network, each camera of the plurality of cameras based on the homography matrix generated for each video frame.
8. The system of claim 7 , wherein identifying the camera motion in each video frame using the optical flow model, comprises:
generating, for each video frame of the plurality of video frames, body pose information for each player in each video frame; and
removing each player from each video frame based on the body pose information.
9. The system of claim 8 , wherein removing each player from each video frame based on the body pose information, comprises:
identifying one or more pixels corresponding to a player in a video frame based on the body pose information; and
removing the identified one or more pixels from the video frame.
10. The system of claim 7 , wherein retrieving the live broadcast video feed for the sporting event comprises:
identifying, from the plurality of video frames, a set of trackable video frames, wherein each trackable video frame in the set of trackable video frames captures a same perspective of the playing surface.
11. The system of claim 7 , wherein generating the homography matrix for each video frame based on the camera motion comprises:
training a further neural network to generate the homography matrix for each video frame based on the subset of labeled video frames and flow fields associated with the camera motion between successive video frames.
12. The system of claim 7 , wherein calibrating each camera based on the homography matrix generated for each video frame comprises:
projecting player coordinates in each video frame to real-world coordinates based on the homography matrix of each video frame.
13. A non-transitory computer readable medium including one or more sequences of instructions that, when executed by one or more processors, perform one or more operations comprising:
identifying, by a computing system, a live broadcast video feed for a sporting event, the live broadcast video feed comprising a plurality of video frames captured by a plurality of cameras remotely located in a sporting venue;
labeling, by the computing system via a neural network, components of a playing surface captured in each video frame to generate a set of labeled video frames;
matching, by the computing system, the set of labeled video frames to a set of templates, wherein each template of the set of templates represents a different camera perspective of the playing surface;
removing, by the computing system, at least one labeled video frame from the set of labeled video frames to generate a subset of labeled video frames, wherein the at least one labeled video frame does not match any template in the set of templates;
identifying, by the computing system, camera motion in each video frame using an optical flow model;
generating, by the computing system, a homography matrix for each video frame based on the camera motion; and
remotely calibrating, by the computing system over a network, each camera of the plurality of cameras based on the homography matrix generated for each video frame.
14. The non-transitory computer readable medium of claim 13 , wherein identifying, by the computing system, the camera motion in each video frame using the optical flow model, comprises:
generating, for each video frame of the plurality of video frames, body pose information for each player in each video frame; and
removing each player from each video frame based on the body pose information.
15. The non-transitory computer readable medium of claim 14 , wherein removing each player from each video frame based on the body pose information, comprises:
identifying one or more pixels corresponding to a player in a video frame based on the body pose information; and
removing the identified one or more pixels from the video frame.
16. The non-transitory computer readable medium of claim 13 , wherein generating, by the computing system, the homography matrix for each video frame based on the camera motion comprises:
training a further neural network to generate the homography matrix for each video frame based on the subset of labeled video frames and flow fields associated with the camera motion between successive video frames.
17. The non-transitory computer readable medium of claim 13 , wherein calibrating, by the computing system, each camera based on the homography matrix generated for each video frame comprises:
projecting player coordinates in each video frame to real-world coordinates based on the homography matrix of each video frame.