analyzing finish photos -...
TRANSCRIPT
ANALYZING FINISH PHOTOSROY VAN HAL
SUPERVISED BY DIRK MEIJER
LEIDEN INSTITUTE OF ADVANCED COMPUTER SCIENCE
OVERVIEWTrack running is a subsection of the eventsthat exist in “track and field” sports. As in alot of sports, the goal in track running is to bethe first competitor to cover a certain distance.During the start, the participants are lined upand wait for a start signal. When it is given,the race has started and a timer keeps trackof the time that has elapsed. When a competi-tor reaches the finish, which is usually a line,their time has to be determined. The finish lineis captured using a line-scan camera. An ope-rator of the timing system used the camera’srecordings to determine a race time for eachcompetitor. However, this is a manual taskand takes time. The goal of this research is toshorten this time by creating a program thatestimates the finish times automatically.
PHOTO FINISH PHOTOSThe finishing competitors are recorded using a line-scan camera. This camera is aligned with thefinish line. The frames that it can record are only a single pixel wide, hence the name. Each framethat is captured is associated with a time stamp that is provided by the timer. The camera operatesat a high frame rate, allowing very accurate finish times for each of the participants. When theframes are chronologically arranged, a finish photo is created, with the location on the finish lineon the vertical axis, and the time on the horizontal axis. The picture below shows an analyzedfinish photo of a 200m race with the time on the horizontal axis, the location on the finish line onthe vertical axis. The finish time for each participant is indicated by a red line.
After the race has finished, the operator of the timing system uses software to analyze the finishphoto. For each participant, he assigns the race time. The time of a finishing competitor is deter-mined by the moment that they touch the vertical plane of the finish line with any part of theirtorso, excluding the head, arms and legs.
RESEARCH QUESTIONThe question that we want to answer is: Is itpossible, and in what extend, to automate thedetermination of track running race times byanalyzing a finish photo using image mani-pulation techniques?
To reduce the delay between finish time andpublishing results, we try to develop a com-puter program that can analyze a finish phototo estimate the finish times of the competitors,to give them an idea of their placements andrace time. The program has to:
• Recognize the torsos of the competitors.The torso of each participant has to berecognized to assign a time stamp.
• Recognize each competitor’s lane. Thisallows identification of the participantsand makes it possible to assign the timesto a name.
• Estimate finish times accurately. The ti-mes that are estimated by the programshould be close to the official time thatthe operator will assign.
• Generate the estimations quickly. Theprogram is not useful if it is slower thanmanually determining the finish times.
RELEVANT WORKJorgensen and Im have proposed an idea ofusing a personal device to determine finishtimes. In their paper, they describe an algo-rithm to detect objects passing a line in a vi-deo. The algorithm chooses and stabilizes theframe in which the object reaches the line. Thisframe corresponds to a specific time stamp.However, the accuracy of this algorithm can-not get close to those of an official timing sy-stem, since these systems camera’s are alignedprofessionally, provide a much higher framerate and have a more accurate timer. Also, of-ficial timing systems are operated by judgeswho have received special training.Jorgensen and Im (n.d.). Fully Automatic Timing for a Personal Device. Stanford University.
APPROACHThe program that will be developed consistsof multiple steps that are executed in order.These steps are shown systematically in the fi-gure below, starting from the top left.
First of all, the finish photo is imported intothe program. Then, a background is computedto separate the finish line and the participantsinto layers. The participant layer is then en-hanced and posture recognition is applied todetect the torsos. These are used to determinea time stamp on which they reached the finishline. Each of these race times then will be lin-ked to a lane in which the competitor has finis-hed, thereby assigning the times to the compe-titors. Finally, these lanes and the correspon-ding race times are shown, and an image thatshows how the times are assigned by the pro-gram and to which lane. The operator can cor-rect the estimations manually if desired.
DETECTING PARTICIPANTSBefore the finish times can be detected, theparticipants need to be recognized first. Theprocess of separating foreground and back-ground is called segmentation. By detectingwhether a pixel is foreground of background,it is compared to the most common pixels inthe picture. The result is an image that showswhich parts are foreground and which arebackground, like in the image below.
The white area’s represent foreground, theblack parts of the image are detected as back-ground. The image has been enhanced usingthe morphological operations of dilation anderosion, which expand and shrink the whitearea’s to counter any noise that was createdduring segmentation. This image can be usedto recognize the torsos of the participants.Then, the finish times can be assigned.