Can the biomechanics of a Horses’ walk predict their ability?
Over the winter we have been working on an extensive project to incorporate biomechanics and machine learning into our assessment criteria when profiling horses at the sales.
In this piece, I will share one such analysis technique utilizing biomechanics to plot the trajectory of a horse as they walk in a pre-sales video.
Our goal is to bring a scientific approach to the traditional art of evaluating the walk of a horse, thus being able to put a quantifiable metric against the statement “that horse is a good walker”.
In order to start the analysis we first had to collect 1000s of videos of horses walking from all over the world. This meant combing the sales catalogues from Tatts, Arqana, Goffs, Fasig-Tipton, Inglis, Magic Millions in order to build up a database of videos. These videos are then mapped to the actual racing results that the horses achieved on the track, using Official Rating where available or RPR where not available as a proxy (converted to OR using a mapping table).
We want to maintain the integrity of the walk videos that we have collected and have separate out those that are Breeze Up Horses / 2yos in Training vs. Yearling videos. Currently this gives us two datasets of ~1500 breeze up horses and ~3000 yearlings to analyze.
In order to map the biomechanic movements of the horse, we first have to train an image recognition model, using a Tensorflow algorithm to recognize and map body parts of a horse as they move through a video. This is developed using an open source software (Deeplabcut), which extracts single images from the the video of a horse walking.
I estimate that over the winter I've spent nearly 250 hours manually tagging body parts against images of horses in order to train this image recognition model.
Notice on the image that we are tagging key body parts of the horse, this is what we will track in the video later on. For example, on each leg we are labelling the hock, the fetlock and the foot.
After labelling thousands of images (I labelled 5000) we are then able to train a model to recognise body parts and the movements of horses as they walk.
After analysis of videos the model is able to produce a mapping like displayed in the video below.
We can then generate a plot of the horse walking over the course of the video and the subsequent movement of each of the body parts. When running over our database of videos of horses walking, this gives us a distinct plot for each horse that we can tag against the racing performance that they achieved (in OR). Thus, building up a picture for us of the trajectories of “good walkers” vs. “bad walkers”
Take a look at some of the examples below.
This is a 100+ rated Vadamos colt, look at the consistent lines of movement on the feet, fetlocks and knees. In graphical format, the colt displays a “smooth walk”.
This is a 100+ rated Kodiac, the profile is fairly similar:
In comparison, this is a 48 rated Churchill colt. Even graphically you can see from the plot that the walk of the horse is not as smooth as the other examples, it’s much more jerky and inconsistent. (Focus on the bottom 4 lines)
This is a 40 rated Filly by Twilight Son. Notice how the movement of the feet, fetlocks and knees has an irregular cadence and doesn’t display the same consistency as the high performing examples above.
After compiling a database of all these type of images, we have trained a further ML model using an Image Classification algorithm, teaching it which images map to high performance track results. Now built we can just upload new images one by one and receive a probabilistic prediction of the horse achieving a certain Official Rating threshold.
If you find that interesting, there’s more to come.
Tom Wilson is the founder of racing and bloodstock analytics start up racing².
racingsquared@gmail.com
