Alright, so today I’m gonna walk you through this little experiment I did with some soccer footage – seeing if I could actually, you know, spot players faking injuries. It was more of a “can I build something that sorta kinda works” kinda deal, not some super serious scientific study.
First things first: data. I spent way too long digging around YouTube for clips of soccer games, specifically looking for instances where players went down holding their leg or something, looking all dramatic. The key was variety – different angles, different players, you name it. I ended up with a messy pile of video snippets, maybe around 50 or so. Total time spent on this part? Probably a couple of hours, easy.
Next up: labeling. This was the most boring part, hands down. I had to go through each clip and basically tag what was happening. Was the player actually tackled? Did they just trip? Or did they just decide to fall down for no reason? I ended up with three categories: “Genuine Injury,” “Possible Foul,” and “Clear Dive.” This was super subjective, I’m not gonna lie. I tried to be as objective as possible, but sometimes it was just a gut feeling. Used a simple video editing tool to mark start and end times of the “injury” event and added those labels. Took like, three or four hours, maybe broken into two sessions because I was losing my mind.
Now for the fun part: the model. I figured I’d try to keep it simple. I didn’t want to build some crazy complex AI thing. I went with a pretty basic convolutional neural network (CNN) using TensorFlow. Why CNN? Well, it’s good at spotting patterns in images, and video is just a series of images, right? I chopped each clip into individual frames – basically a picture for every fraction of a second. Then, I fed those frames into the CNN. The idea was to train the model to recognize the difference between the body language and movements associated with a real injury versus a fake one.
Training Time! This is where the computer does all the work (hopefully). I split my labeled data into training and testing sets. The training set is what the model learns from, and the testing set is what I use to see how well it learned. I messed around with the model’s settings – things like the number of layers, the learning rate, stuff like that. Honestly, a lot of it was just trial and error. I’d run the training, see how it performed on the test set, tweak the settings, and repeat. This process took, I dunno, maybe a whole afternoon. The training itself took a couple of hours each time I ran it.
The Results? Okay, so here’s the thing: it wasn’t perfect. Far from it. The model could pretty reliably spot a genuine tackle. Those were usually pretty obvious – player gets knocked to the ground, writhing in pain. The “Possible Foul” category was hit-or-miss. Sometimes it got it right, sometimes it didn’t. But the “Clear Dive” category? That’s where it struggled. I think the subtle differences in body language were just too nuanced for the model to pick up consistently. It would get maybe 60-70% accuracy on the test set, which isn’t terrible, but not exactly groundbreaking.
What I Learned. This whole thing was a good reminder that even simple AI projects can be a lot of work. Also, data quality is HUGE. My dataset was small and probably not super accurate to begin with, so that definitely limited the model’s performance. If I were to do it again, I’d spend way more time collecting and labeling data, and maybe try a more sophisticated model. But hey, it was a fun experiment, and I learned a ton along the way.
Next Steps? Honestly, probably nothing. I’ve got other projects I want to work on. But if someone else wants to take this and run with it, be my guest! Just remember to give me a shout-out when you build the definitive “fake injury detector.”