Over the past few years, we have live tracked thousands of athletes. There have been 2 big questions from our customers:
There are several factors that complicate position triangulation and course progress. GPS coordinates can occasionally be inaccurate, and the course itself may overlap the same locations many times or has gaps in cellular coverage. In addition, the course may be in a relatively small space which makes disambiguation of exactly which part of the course that GPS coordinate should correlate with difficult. It is a very hard problem to solve.
The great news is we’ve accomplished a new level of accuracy and reliability using our AI framework. And we’re able to do this along with the difficult complications above. This framework will serve as the foundation for which we can build an even better tracking experience for athletes and their fans, and much more!
Knowing where someone is on the course is the first, most important bit of information we need. It unlocks features you’ll love and have been asking for. Things like knowing when your athlete is finishing or their current rankings among their peers become possible. We‘ll be able to better identify anomalous patterns in race progress such as going off course, stopped moving for too long, going in the wrong direction, or perhaps everyone in the event is slowing due to weather. An accurate ETA helps you meet your athlete anywhere on the course, alert or notify you when they are about to arrive or have stopped for too long. All of these things and more have been filed in our patent application. More importantly, as the features are developed they will progressively be made available to you throughout this year.
At first glance, one may think the above questions are trivial to answer. However, in practice they are quite challenging for any GPS tracking system. In order to treat them with proper respect, let us take a tour through the world of GPS tracking and consider the various complexities. If you have used our platform in the past then you are already familiar with how BAT works. Otherwise here is a quick screenshot of our service in action:
1. Multiple Loops: In the image above, we see an athlete with Id 2025 moving along the blue bike segment of the Ironman 70.3 triathlon race in Galveston, Texas on April 8th, 2018. The location data is displayed in real time, allowing friends and family to follow along and have peace of mind that their athlete is safe. You may have noticed, however, that this view does not have any information pertaining to our problems of interest. While our map includes mile markers throughout the race course, a segment may contain multiple loops through the same paths.
For example, mile markers 10, 15, 40, and 45 are all on the same line. Given this image alone, one may guess that the athlete is somewhere between mile 10 and 15 due to the south-west direction indicated by the arrow. But how do we know if this is the first time an athlete reaches this location, or the second time, or the third time, and so on? Depending on the speed, it could be the first for one athlete but the second for another. In mathematical terms, reliably determining the exact loop from just this GPS location is an ill-defined problem with ambiguous solutions.
2. Multiple Out-and-Backs: To make the problem harder, let's view another example. Below is the actual run course of the same event where athletes must run 3 loops through all of the red paths to finish the segment. As before, the athlete's location is shown by the black pin. This particular location is interesting because it lies on an out-and-back section. Thus, through the entire segment, each athlete will pass by this point at 6 different times! If one were to take a random guess, that would equate to only 16.7% chance of winning.
3. Location Deviation: While GPS signals are accurate to several meters most of the time, occasionally they could be erroneous depending on the local environment such as buildings, trees, and other obstacles. Additionally, as athletes run through the race, it is normal to deviate slightly from the preset paths. Due to these conditions, the resulting position may be more difficult to interpret. Let's examine this case through another example below, still in the same Ironman 70.3 Texas run course. Since this reported location is at the center of 3 surrounding paths, it's possible that the athlete could actually be at any of the 3 blue circles. And since there are 3 loops in total, each blue circle is visited 3 different times, thus making it a total of 9 different possibilities. The chance of guessing it right is now only 11%.
4. Cellular Coverage: To further complicate the problem, coverage can be spotty in some areas leading to extended service outage. From this perspective, Ironman Lake Placid is one of the most challenging events in North America. Although BAT will offer 2 options in the future to increase coverage, our service must be able to handle these situations regardless. The following image from Lake Placid 2017 event shows 2 successive signals reported nearly ~1.5 hours apart due to lack of cellular connection. Since the upper right location contains an out-and-back section, we must choose between the "out" vs the "back" points. If the direction arrow is accurate, the problem could be easy. If not, however, the extended outage time could again lead to ambiguous triangulation.
Despite these obstacles, BAT can now reliably determine the precise location of an athlete on each segment of the course, regardless of exogenous conditions and using only GPS signals. By closely observing and monitoring each athlete's performance, BAT constructs a Machine Learning model that is personalized to each athlete and makes predictions accordingly. Each model captures the statistical distribution of movement speeds and past data to perform highly accurate triangulation. Future improvements are possible via integration with checkpoint-based timing systems such as RaceResult's TrackBox. Here're a couple videos to demonstrate these predictions on actual race data (note: select 1080p quality if your video is blurry):
The first video is on the challenging run course of Ironman 70.3 Texas 2018. The second video is from Ironman Santa Rosa 2017 and shows normal biking condition with a small lapse in coverage starting at 00:21 sec. Finally, in the third video, there is a large gap of coverage lasting for nearly 2 hours as seen at the beginning. For all 3 different races with varying environments, however, our predictions are still accurate.
You may have also noticed another feature from these videos: the ability to replay your past races. Very soon every athlete who has used our platform will be able relive their past events and share those moments with others. We hope these features will make your live tracking experience better and more connected to the people you care about. Here's to your next successful event!
* We would like to thank Long Dao, Cyndi Wells, Liisa Travis and Jody Ferrell for the permissions to use their race data throughout this post. All images, screenshots, and videos are produced based on actual data. Videos were recorded and played back at faster than actual speeds.
BAT Crew - Luong, Christy & Seth