Biomechanics-Guided Standard Action Modeling and Lightweight ST-GCN for Interpretable Phase-Based Action Quality Assessment

Imagine you are watching a friend try to teach a robot how to play basketball. If you just tell the robot, "You missed the shot," it learns nothing. But if you say, "Your elbow was too low, you didn't bend your knees enough, and you let go too early," the robot can actually fix its game. This is the heart of a field called Action Quality Assessment (AQA). For years, computers have been great at recognizing what action is happening (like "jumping" or "running"), but they have struggled to explain how well it was done. Traditional methods often require expensive, bulky cameras and sensors that look like a sci-fi movie set, making them impossible to use in a regular school gym. The big question researchers are trying to solve is: Can we use a simple video from a phone or a standard camera to not only give a score but also explain exactly where a movement went wrong, all without needing a lab full of equipment?

This paper introduces a clever new system designed to grade the standing long jump—a classic test of explosive power—using just a video. The researchers, led by a team from Capital Normal University, built a digital "coach" that breaks the jump down into its tiny, mechanical parts. Instead of just looking at the final distance jumped, their system watches the athlete like a biomechanical detective. They created a "standard template" based on expert knowledge, which acts like a perfect ghost of an ideal jump. Then, they used a lightweight AI model (a type of brain called a ST-GCN) to compare the real athlete's video against this perfect ghost. The result isn't just a number; it's a detailed report card that says, "Your take-off angle was too low," or "You didn't bend your knees enough when you landed." The team tested this on 200 people and found their AI could grade the jumps with a level of accuracy that matches human experts, but it does it with a tiny fraction of the computing power, making it possible to run on regular computers in schools.

The Digital Coach and the Perfect Ghost

Think of the standing long jump as a complex dance with five distinct moves: getting ready, crouching down, exploding upward, flying through the air, and landing safely. In the past, if a coach wanted to grade this, they would have to watch the video frame-by-frame, guess the angles, and write down notes. It's slow, and two coaches might disagree.

The authors of this paper decided to build a digital coach that never gets tired and never disagrees with itself. First, they had to teach the computer how to see. They took videos of 200 participants (100 men and 100 women) jumping and used a tool called MMPose to find the "skeleton" of the person in the video. It's like drawing a stick figure over the person in every single frame of the video. Then, they used another tool, VideoPose3D, to lift that stick figure into 3D space, so the computer understands depth, not just a flat picture.

But seeing the skeleton isn't enough; the computer needs to know what a good jump looks like. So, the team built a Standard Action Template. Imagine this as a "Perfect Ghost" jump. This ghost doesn't have a body, but it has a set of rules: "The knees should bend to 80–100 degrees," "The take-off angle should be between 19 and 40 degrees," and "The landing should be soft." This ghost was created by combining scientific studies with the knowledge of real sports experts.

The Lightweight Brain

Now comes the star of the show: the Lightweight ST-GCN. In the world of AI, "GCN" stands for Graph Convolutional Network, which is a fancy way of saying a computer brain that understands how body parts are connected (like how your knee moves when your hip moves). Usually, these brains are huge, heavy, and slow, requiring powerful supercomputers to run.

The researchers wanted something that could run on a regular computer, maybe even a laptop in a school gym. So, they built a lightweight version. Think of it like shrinking a massive, complex library down to a single, perfectly organized notebook. They cut out unnecessary parts, removed complex attention mechanisms (which are like the brain trying to look at too many things at once), and simplified the structure.

The result is a model that is incredibly small. It has only 0.48 million parameters (the "memory" of the brain) and requires 2.7 GFLOPs of computing power. To put that in perspective, a standard, heavy-duty version of this brain would need 3.10 million parameters and 16.4 GFLOPs. The new model is about 84.5% smaller and much faster. It can analyze a jump in just 18.2 milliseconds on a high-end computer, or 64.3 milliseconds on a standard CPU. That's fast enough to give feedback almost instantly, like a coach shouting advice while you are still in the air.

How It Grades and Explains

Here is where the magic happens. The system doesn't just look at the whole jump and guess a score. It breaks the jump into three main scoring zones: Take-off, Flight, and Landing.

1. The Take-off Zone: This includes the preparation, the crouch, and the explosion. The AI checks if the athlete bent their knees enough and if they swung their arms correctly.
2. The Flight Zone: This is the time in the air. The AI checks if the body stayed stable and if the legs were pulled back properly.
3. The Landing Zone: This is the impact. The AI checks if the knees bent to cushion the fall and if the body stayed balanced.

The AI gives a score for each zone and then combines them into a total score. But the best part is the feedback. If the "Take-off" score is low, the system looks at the "Perfect Ghost" template and asks, "What was different?"

• Did the knees not bend enough? -> Feedback: "Insufficient push-off."
• Was the jump too flat? -> Feedback: "Low take-off angle."
• Did the arms stay still? -> Feedback: "Insufficient arm swing."

The Results: Does It Work?

The team tested their digital coach against three human experts. The results were impressive. The AI's scores matched the human experts with a correlation of 0.94 (where 1.0 is a perfect match). In terms of error, the AI was off by only 0.12 points on average (RMSE), compared to 0.33 points for the human experts and 0.26 points for other computer methods.

The paper suggests that this method is particularly good at the take-off phase, where it was most consistent. The flight and landing phases were slightly harder to judge perfectly, likely because the body moves so fast and the camera can sometimes lose track of the joints for a split second. However, even with these tiny hiccups, the system was still more accurate than traditional computer methods and just as good as the human experts.

The researchers also compared their "split-phase" method (grading the three parts separately) against a method that just looked at the whole jump at once. The split-phase method won hands down. It's like grading a math test by checking the addition, the multiplication, and the final answer separately, rather than just looking at the final number. This approach helped the AI spot small mistakes that would have been hidden if it only looked at the total score.

The Bottom Line

This paper suggests that we don't need expensive labs to get professional-level feedback on sports movements. By combining a "Perfect Ghost" template with a tiny, fast AI brain, we can turn a simple video into a detailed coaching session. The system is fast, accurate, and, most importantly, interpretable—it tells you why you got a certain score, not just what the score is.

The authors are careful to note that this isn't a replacement for a real human coach or a lab with high-tech sensors. It's a tool for schools and grassroots training where resources are limited. They also admit that their test group was mostly young adults, so it might need more testing on kids or older athletes. But for now, they have shown that with a little bit of clever engineering, a computer can learn to watch a jump, understand the physics, and give you a high-five or a gentle nudge to try again.