<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Mohammed Ebrahim Alzaabi on Medium]]></title>
        <description><![CDATA[Stories by Mohammed Ebrahim Alzaabi on Medium]]></description>
        <link>https://medium.com/@m.ebrahim.alzaabi?source=rss-7ad183266eb7------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*vcJAAXQlF9JC1KjNADio1w.jpeg</url>
            <title>Stories by Mohammed Ebrahim Alzaabi on Medium</title>
            <link>https://medium.com/@m.ebrahim.alzaabi?source=rss-7ad183266eb7------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 07:38:14 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@m.ebrahim.alzaabi/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Can a Webcam Tell a Good Push-Up From a Bad One? I Built One to Find Out]]></title>
            <link>https://medium.com/@m.ebrahim.alzaabi/can-a-webcam-tell-a-good-push-up-from-a-bad-one-i-built-one-to-find-out-8d047db9df9c?source=rss-7ad183266eb7------2</link>
            <guid isPermaLink="false">https://medium.com/p/8d047db9df9c</guid>
            <category><![CDATA[fitness]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[calisthenics]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Mohammed Ebrahim Alzaabi]]></dc:creator>
            <pubDate>Wed, 04 Nov 2026 14:51:18 GMT</pubDate>
            <atom:updated>2026-11-04T14:51:18.586Z</atom:updated>
            <content:encoded><![CDATA[<p><em>How I turned Google’s pose model into a push-up counter that ignores half reps and catches hips that sag or rise too high, and what broke along the way.</em></p><p>I’m Mohammed, a 12th-grade student in Ras Al Khaimah, and I train on my own. No coach, no training partner, just me trying to copy what I see in videos. A couple of months ago, that caught up with me. I hurt my knee doing a one-leg squat, and I still don’t know exactly what I did wrong, because nobody was watching my form.</p><p>After that, one question stuck with me. What if my camera could watch me instead? In my <a href="https://medium.com/@m.ebrahim.alzaabi/your-camera-as-a-coach-how-ai-pose-estimation-could-help-beginners-train-with-correct-form-325d17dc24f6">Learn article</a>, I looked at <strong>pose estimation</strong>, AI that finds a person’s joints in an ordinary photo or video. On paper, it looked like it could have warned me before I got hurt. But reading about something and building it are two very different things.</p><p>So this month, for my TKS project, I stopped reading and started building. I gave myself one exercise and one question. Can a normal webcam tell a good push-up from a bad one? I picked push-ups because most beginners start with them, and one of the most common mistakes, hips sagging toward the floor, is easy to see from the side.</p><p>This is step two of a bigger plan. I’m building an AI coach for people who train alone, one piece at a time, and I’m writing about every step here, including the parts that break. This month is the foundation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/720/1*7pk3Anz44jj78ZSwuDhunA.gif" /><figcaption>The app counts the rep and judges the form as it happens, from nothing but a normal video. Source: my app running on a free Pexels video.</figcaption></figure><h4>Summary</h4><ul><li><strong>What it does:</strong> it watches you do push-ups from the side, counts only full reps, and warns you when your hips sag or rise too high.</li><li><strong>How:</strong> Google’s free MediaPipe model finds 33 points on your body. The app turns those points into two angles, one at the elbow and one along the body, and uses them to count reps and check your form.</li><li><strong>Results:</strong> in my tests, it counted all 15 full push-ups and none of the 5 half reps. It flagged all 5 reps where I let my hips sag on purpose, 4 more where I only let them sag slightly, and all 5 where my hips were too high. After I tuned it, it matched my own labels on all 10 good reps.</li><li><strong>Code:</strong> free to read and run at <a href="https://github.com/mohammed-ebrahim-code/pose-coach">github.com/mohammed-ebrahim-code/pose-coach</a></li></ul><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fa5r4rERfAKU%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Da5r4rERfAKU&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fa5r4rERfAKU%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/7e4c7023fa95e00aa3a835dff3239fea/href">https://medium.com/media/7e4c7023fa95e00aa3a835dff3239fea/href</a></iframe><h3>How it works</h3><p>A video is just a series of still images called frames. The app takes each frame, about 25 times a second, and does the same five things.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jQQjE4sFCdtAAqyp5t1fbw.png" /><figcaption>Every frame goes through the same five steps, explained one by one below. Source: diagram made for this article.</figcaption></figure><h4>Step 1: Find the joints</h4><p>I use Google’s <strong>MediaPipe Pose Landmarker</strong>. For every frame, it returns 33 points, called landmarks, such as the shoulders, elbows, hips and ankles. Each point has a position (x and y) and a <strong>visibility</strong> score from 0 to 1, which says how sure the model is that the point can actually be seen.</p><p>The app ignores any point with a visibility below 0.5. I chose that because of my Learn-article test, where joints hidden behind my body were the main reason angles came out wrong. For a form checker, skipping a frame is better than giving someone wrong feedback.</p><p>From the side, one half of the body is hidden, so the app uses the side the model can see better and uses that <strong>same side</strong> for both the arm and the body.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LzLpqEApsQAMMs3Skn9BjQ.jpeg" /><figcaption>The model finds the joints, and the app draws them as a skeleton. Blue is the arm it measures, and green is the body line it checks. Source: my app running on a free Pexels video.</figcaption></figure><h4>Step 2: Measure two angles</h4><p>Points on their own don’t tell you anything about a push-up, but the angles between them do. The app uses two of them.</p><ul><li>The <strong>elbow angle</strong> (shoulder, elbow, wrist) is around 170° with straight arms and drops below 90° at the bottom of a push-up. The app uses it to <strong>count</strong> reps.</li><li>The <strong>body angle</strong> (shoulder, hip, ankle) is close to 180° when your body is a straight line, and gets smaller when your hips sag or rise too high. The app uses it to <strong>check form</strong>.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F70EdDjFmjiDAJi4C3B6eQ.png" /><figcaption>One angle counts the reps, and the other checks the form. Source: diagram made for this article.</figcaption></figure><p>To get the angle at a joint, I draw two arrows that both start at the joint, one to each neighbouring point. The <strong>dot product</strong> is a calculation that measures how much two arrows point in the same direction. Divide it by the lengths of the two arrows, take the arccos (the inverse of cosine), and you get the angle between them.</p><pre>θ = arccos( (BA · BC) ÷ (|BA| × |BC|) )</pre><p>Here, B is the joint (for example, the elbow), A and C are the points on either side, BA and BC are the two arrows, and |BA| means the length of arrow BA.</p><p>One detail matters a lot. MediaPipe gives each position as a fraction of the image, from 0 to 1, so 0.5 means halfway across. My frames are wider than they are tall, so moving 0.1 sideways covers more of the real picture than moving 0.1 up or down. Using those numbers directly would change every angle, so the app first converts every point to pixels. This is the kind of detail I never thought about in my Learn article, and it only showed up once I was writing the code.</p><h4>Step 3: Count reps</h4><p>The app only needs to remember one thing between frames, whether you are <strong>up</strong> or <strong>down</strong>. When the elbow angle drops below <strong>90°</strong>, you’re down. When it comes back above <strong>160°</strong>, you’re up again, and that’s the moment the rep counts. This is called a <strong>state machine</strong>, which means the program is always in one state, and only a specific event moves it to the other.</p><p>I use two separate lines, 90° for down and 160° for up, instead of one line for both. When I tested pose estimation for my Learn article, the angle moved by about ±2° even when I held completely still. With only one line, for example at 120°, those small movements would cross it again and again and count fake reps. With 70° between the two lines, that can’t happen. A half rep also never gets below 90°, so it never counts.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kK51us_vvqtzI9q7tBDUnw.png" /><figcaption>Only a full rep crosses both lines, so the half rep on the right is never counted. Source: diagram made for this article.</figcaption></figure><h4>Step 4: Judge the rep</h4><p>Counting reps was the easy part, because a rep is clearly defined by the elbow angle. Judging form is harder, because I had to decide what “good” means and turn it into a number. For this replicate, I focused on one thing, the hips. Are they sagging toward the floor, or piking up too high?</p><p>During each rep, the app keeps track of the <strong>lowest body angle</strong> it has seen. When the rep finishes, it checks that one number. Below <strong>155°</strong> means the hips weren’t straight. The app then checks which way they went, sagging or piking. (I explain how I chose 155°, and how it tells the two apart, in the testing section below.)</p><p>Judging the whole rep at once gives one clear answer per rep, instead of a message that flickers on and off every frame. Because it keeps the lowest angle, it also catches a sag that only happens for a moment at the bottom.</p><p>The app only starts watching the body once the elbow is clearly bent (below <strong>130°</strong>). Before that, you might still be getting into position, and that shouldn’t count against your first rep.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F_3VeKutN6n-hdXCR4_joA.jpeg" /><figcaption>The app judges a rep only after it ends. Here the lowest body angle stayed above 155°, so it shows GOOD REP. Source: my app running on a free Pexels video.</figcaption></figure><h4>Step 5: Show it on screen</h4><p>The screen shows the rep count, the good and bad reps, and whether you’re up or down. On the right, a bar shows the elbow angle live, with lines at 90° and 160°, so you can see exactly when a rep will count. After each rep, a “GOOD REP”, “HIPS SAGGING” or “HIPS TOO HIGH” message appears for one second. If your arm or body isn’t fully in view, the app asks you to show your whole body instead of guessing.</p><p>On a webcam, the image is flipped so it feels like looking in a mirror, but only <strong>after</strong> the joints are found. If you flip it first, MediaPipe thinks your right arm is your left. This was a bug I ran into myself.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eSp-lHRLpGkzNzqgSYJ2bQ.jpeg" /><figcaption>The bar on the right shows the rep counter working. At the bottom of the rep, the elbow is at 57°, below the 90° line, so the app switches to DOWN. Source: my app running on a free Pexels video.</figcaption></figure><h3>What I built on, and what I added</h3><p>I didn’t follow a tutorial or copy someone else’s project. I started from Google’s official <a href="https://developers.google.com/edge/mediapipe/solutions/vision/pose_landmarker/python">Pose Landmarker guide for Python</a>, and when I got stuck, I asked an AI assistant (Claude) for explanations of things that were new to me. AI also helped me put the demo video together.</p><p>From the guide, I took the basics of running the model. That means loading it, running it on a video frame by frame with a timestamp, and knowing which number belongs to which of the 33 points. The angle formula is standard vector maths.</p><p>Everything that turns those 33 points into a push-up checker, I added myself.</p><ul><li><strong>The two-line rep counter</strong> (90° down, 160° up), so small jitters can’t fake a rep and half reps never count</li><li><strong>Judging each rep once,</strong> at the end, from its lowest body angle</li><li><strong>The 130° rule,</strong> which only starts judging the body once the elbow is bent</li><li><strong>Converting points to pixels</strong> before measuring angles, so wide frames don’t change them</li><li><strong>The visibility filter,</strong> and using the better-seen side for both the arm and the body</li><li><strong>Flipping the image only after</strong> the joints are found</li><li><strong>Choosing 155°</strong> from my own test data</li><li><strong>Telling sagging from piking,</strong> by checking whether the hip is above or below the line from shoulder to ankle (an idea that came from my director’s feedback)</li></ul><p>These choices are my real contribution, and several of them came from something breaking.</p><h3>What broke, and what it taught me</h3><p>Most of what I learned came from things going wrong. Looking back, I understood what most of my lines did, but not always the reason behind them. Knowing why each part was needed, and testing each function on its own, would have made finding these bugs much faster. These are the three that taught me the most.</p><h4>The angle that came out backwards</h4><p>The first test of my angle function gave 0° and 135° where I expected 180° and 45°. Every wrong answer was exactly 180° minus the right one. That pattern was the clue. One arrow was pointing the wrong way, because I had written b - a instead of a - b. Both arrows have to start at the joint.</p><h4>The bug my tests couldn’t see</h4><p>In the formula, you divide by both lengths, |BA| × |BC|. I forgot the brackets, so Python divided by the first length and then multiplied by the second. My tests still passed, because every test point I used was exactly 1 unit away from the joint, and multiplying or dividing by 1 changes nothing. The lesson is to pick tests that can actually expose a bug.</p><h4>The bug only someone else could find</h4><p>When I ran the app on the Pexels video, the man’s first rep was marked bad even though his form was good. Before starting, he got into position with his hips raised and his arms not fully straight, and my app was already judging his body during that setup. That’s where the 130° rule in Step 4 came from. Testing on someone else’s body found a problem that testing on myself never would have.</p><h3>Testing it</h3><p>A demo only shows that something <em>can</em> work. To find out how well it works, I tested it on purpose with several sets of push-ups. The app saves every rep to a file (good or bad, plus its lowest body angle), so these numbers come straight from that log.</p><p>One thing to keep in mind is that I labelled every rep myself, as good, half, sagging or piking. So when I say the app got a rep right, I mean it matched my own label. Nobody else, like a coach, checked them.</p><h4>My own push-ups</h4><p>I put my laptop camera on the floor, filming from the side with my whole body in frame. The app ran at about 25 frames per second.</p><ul><li><strong>Set 1, 10 good push-ups:</strong> all 10 counted. My first threshold (160°) marked 2 of them as bad. After I changed it to 155° (explained below), it matched my labels on <strong>all 10</strong>.</li><li><strong>Set 2, 5 half reps:</strong> none counted. But these half reps stopped well above 90°, so they were an easy case. Reps that stop just short of the line, around 95–110°, would be a harder test, and I haven’t tried those yet.</li><li><strong>Set 3, 5 push-ups with my hips sagging on purpose:</strong> all 5 counted, and all 5 flagged as sagging.</li></ul><h4>Slight sags</h4><p>The sags in Set 3 were on purpose and exaggerated, so they were easy to catch. The hard cases are slight sags, which land in the gap between my good and sagging reps, from 150° to 160°. So I did four more reps where I only let my hips drop a little. Their lowest body angles were 150°, 144°, 149° and 153°, and the app flagged all four as sagging. Two of them, at 150° and 153°, were inside that gap.</p><p>That’s a good sign, but the margin is small. A sag that only reaches 156° to 159° would still pass as a good rep, so very slight sags are where the app is weakest.</p><h4>Sagging vs piking</h4><p>Piking is the opposite mistake, where the hips go too high. At first, the app couldn’t tell the two apart, because hips too low and hips too high both make the body angle smaller. It could only say “hips not straight”, and that was one of the limitations I cared about most, because a coach doesn’t just say “wrong”, they tell you which way to fix it.</p><p>The fix turned out to be small. If the hip is below the straight line from shoulder to ankle, the hips sagged. If it’s above the line, they piked. I added this check and did 5 push-ups with my hips too high, and the app flagged all 5 as piking.</p><h4>Testing on someone else</h4><p>I also ran it on a free side-view push-up video from <a href="https://www.pexels.com/video/4742661/">Pexels</a>. It counted all 4 of his reps and judged all 4 as good form. That’s also the test that found the setup bug described above.</p><p>These are small tests, on one person and one video, so they show that the app works for me. They don’t yet show that it works for everyone.</p><h4>How I chose 155°</h4><p>To find the number that separates “straight” from “sagging”, I recorded good push-ups and push-ups with my hips sagging on purpose, and looked at the lowest body angle of each rep.</p><ul><li><strong>Good reps:</strong> 160°, 163°, 163°, 164°, 165°</li><li><strong>Sagging reps:</strong> 143°, 145°, 149°, 150°</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F_XJ91tdl3E-p5AdmLPp4g.png" /><figcaption>Each dot is one rep. My good and sagging reps form two groups with a gap between them, so I put the line in the middle of the gap. Source: my own test data.</figcaption></figure><p>The two groups never overlap. Good reps stayed at 160° or above, and sagging reps stayed at 150° or below. My first threshold, 160°, sat right on the edge of the good reps, so a good rep that dipped even slightly was marked bad, which is what happened twice in Set 1. Putting the line at <strong>155°</strong>, in the middle of the gap, leaves room on both sides.</p><p>Even my good reps only reached 160–165°, not the 180° of a perfectly straight line. That’s because I was wearing loose clothes, which make it harder for the model to place the hip exactly. It’s also exactly why the threshold has to come from real data and not from a guess.</p><h3>Limitations</h3><ul><li><strong>Side view only.</strong> The angles are measured flat on the image (in 2D), so they only work when the movement happens across the screen. When I pointed my arm toward the camera, the measured angle lagged behind my real arm. MediaPipe also gives 3D “world landmarks”, positions in metres measured from the middle of the hips, which might fix this because they don’t depend on where the camera is. I haven’t tried them yet.</li><li><strong>Tuned on one person.</strong> The thresholds come from me, in one room, in loose clothes. Other bodies, cameras and lighting may need different values.</li><li><strong>The whole body must be visible.</strong> If the ankles drop out of view, the app pauses instead of guessing. Sometimes my clothes covered my ankles, and that stopped the whole app.</li><li><strong>One person, one exercise.</strong> Push-ups only, for now.</li></ul><p>Two of these matter most for my project. The first is that the thresholds and the labels come from only me. Before this can coach anyone else, I need to test it on more people and check its results against a qualified coach. The second is the side view, because it means this app doesn’t yet solve the problem that started it.</p><h4>Back to my knee</h4><p>I hurt my knee doing a one-leg squat, but this app only checks push-ups, from the side. Many knee mistakes, like the knee caving inward, are easiest to see from the front, so even with a squat mode, this version would probably miss them. Getting there needs a front camera view or 3D landmarks, and that’s one of the first things I want to try next.</p><h4>A note on safety</h4><p>This isn’t medical advice. Pose Coach is a student project that gives basic feedback on one exercise, and it can be wrong. It can’t feel pain or see an injury. If something hurts, stop, and if you’re injured, talk to a doctor or a qualified coach.</p><h3>What I learned</h3><p>In my Learn article, I wrote that finding joints is a solved problem but judging form is not. Building this showed me what I was missing. Even finding joints comes with complications, because clothes, lighting and camera angle all changed how I had to design the app. And judging form really is the hard part, since every angle and every threshold in this app is a decision I made from my own testing. The AI gives you the points. Deciding what “good” means is still up to a person.</p><p>So I don’t think AI will replace a knowledgeable coach. But it can give basic, consistent form feedback to people who have no coach at all, like me when I hurt my knee.</p><h3>What’s next</h3><p>The next stage of my project is the iterate. My director suggested starting small, with one question. How can data from a WHOOP, a fitness band that tracks recovery, sleep and strain, make my training better?</p><ul><li><strong>Pull the data.</strong> Connect to WHOOP’s official API and download my recovery, sleep, strain and workout history.</li><li><strong>Analyze it.</strong> Look for patterns, like how my sleep and recovery line up with how hard my workouts are.</li><li><strong>Use it.</strong> Decide what the data should change, like going easier on days when my recovery is low.</li></ul><p>After that, the create brings everything together in one web app, with this form checker, workout plans that use the WHOOP data, and Arabic as well as English. The long-term goal is a full AI coach. You set a goal, it plans your workouts, watches you with this system, explains each mistake, and adapts your next session. I also want to add more exercises (squats, pull-ups, dips, handstands) and, most importantly, compare the app’s results with a qualified coach’s to see how often we agree.</p><p>If you try the code on your own push-ups, I’d love to hear how it does.</p><h3>About me</h3><p>I’m Mohammed Ebrahim AlZaabi, a 12th-grade student in Ras Al Khaimah and a member of The Knowledge Society (TKS). My main focus area is artificial intelligence, and I’m interested in using AI to make people’s lives easier. With this project, I hope to make training more accessible and safer for beginners.</p><h3>Links</h3><ul><li>Code: <a href="https://github.com/mohammed-ebrahim-code/pose-coach">github.com/mohammed-ebrahim-code/pose-coach</a></li><li>Learn article: <a href="https://medium.com/@m.ebrahim.alzaabi/your-camera-as-a-coach-how-ai-pose-estimation-could-help-beginners-train-with-correct-form-325d17dc24f6">Your Camera as a Coach</a></li><li>Demo video: <a href="https://www.youtube.com/watch?v=a5r4rERfAKU">YouTube</a></li><li>Pose model: <a href="https://ai.google.dev/edge/mediapipe/solutions/vision/pose_landmarker">Google MediaPipe Pose Landmarker</a>, and the <a href="https://developers.google.com/edge/mediapipe/solutions/vision/pose_landmarker/python">Python guide I followed</a></li><li>Test video: <a href="https://www.pexels.com/video/4742661/">free push-up video on Pexels</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8d047db9df9c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your Camera as a Coach: How AI Pose Estimation Could Help Beginners Train With Correct Form]]></title>
            <link>https://medium.com/@m.ebrahim.alzaabi/your-camera-as-a-coach-how-ai-pose-estimation-could-help-beginners-train-with-correct-form-325d17dc24f6?source=rss-7ad183266eb7------2</link>
            <guid isPermaLink="false">https://medium.com/p/325d17dc24f6</guid>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[calisthenics]]></category>
            <category><![CDATA[fitness]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Mohammed Ebrahim Alzaabi]]></dc:creator>
            <pubDate>Sun, 04 Oct 2026 04:31:57 GMT</pubDate>
            <atom:updated>2026-10-08T13:36:50.689Z</atom:updated>
            <content:encoded><![CDATA[<p>A couple of weeks ago, I injured the lower part of my knee doing a one-leg squat. I had watched tutorials on the correct form, but I still don’t know exactly what I did wrong. And that’s the problem: when you train alone, no one is watching your form, not even you.</p><p>I’m not the only one. As working out alone becomes more popular, many people train without a coach or any professional supervision, and many get injured before they even realize their form was wrong.</p><p>What if the camera on your phone or laptop could be that missing coach?</p><p><strong>So, can pose estimation actually be trusted to coach beginners?</strong></p><h3>The problem</h3><p>In the UAE, many beginners start strength and bodyweight training on their own, at home or outdoors, with no coach to check their form.</p><p>Data from nearby Saudi Arabia shows how common these injuries can be. In a <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC11297712/">nationwide survey in Saudi Arabia</a>, over half of the people surveyed reported at least one exercise-related injury.</p><p>Beginners are especially at risk. In <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC5315259/">one study</a>, people with under six months of training experience had more than twice as many injuries per training hour as more experienced athletes. The athletes in that study blamed 20% of their injuries on improper technique.</p><p>Yet most beginners have no way to get real-time feedback on their form. And although the Saudi data shows the problem is real nearby, those numbers don’t apply directly to the UAE. I couldn’t find any UAE-specific data on how often these injuries happen, so that remains a gap.</p><h3>What is pose estimation?</h3><h4>Keypoints and confidence scores</h4><p>Pose estimation is a computer vision technology that finds a person in an image or video and locates their <strong>keypoints</strong>: the joints and landmarks of the body, such as the shoulders, elbows, wrists, hips, knees, and ankles. Researchers have worked on this problem for decades, and deep learning has made it far more accurate over the last ten years.</p><p>Different models detect different numbers of keypoints. The standard COCO dataset uses 17. Google’s <a href="https://www.research.google/blog/on-device-real-time-body-pose-tracking-with-mediapipe-blazepose/">BlazePose</a> detects 33, including extra points on the hands and feet that are useful for fitness. Research models like Meta’s <a href="https://github.com/facebookresearch/sapiens">Sapiens</a> go up to 308.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/900/1*jsQg6B25XQ9VUpeQC-C7BQ.jpeg" /><figcaption>The 33 keypoints BlazePose tracks, including the extra points on the hands and feet that the standard 17-point layout doesn’t have. Source: <a href="https://ai.google.dev/edge/mediapipe/solutions/vision/pose_landmarker">Google, MediaPipe Pose Landmarker documentation</a>.</figcaption></figure><p>Most models also give each keypoint a <strong>confidence score</strong> that shows how sure the model is about it. This is extremely useful when a body part is hidden or out of frame, because an app can ignore the points it isn’t sure about.</p><h4>2D vs. 3D</h4><p>There are two main types. <strong>2D</strong> pose estimation places each joint on the flat image (x, y), while <strong>3D</strong> adds depth (z).</p><p>What surprised me most was that BlazePose runs in real time on an ordinary smartphone while staying accurate enough for fitness, which makes it a strong option for live form tracking in an app.</p><h3>How it works</h3><h4>Finding the joints</h4><p>There are two main ways a model can output where a joint is. The first is <strong>direct regression</strong>: the network looks at the features of the image and directly predicts an (x, y) coordinate for each keypoint.</p><p>The second is <strong>heatmaps</strong>: for each keypoint, the model predicts a map of the probability that the keypoint is at each location in the image, and the highest point on the map becomes the joint’s position. <a href="https://doi.org/10.1145/3603618">Heatmaps</a> turned out to be more accurate and stable, which is why most models moved to them. The trade-off is that they are heavier to compute, because the model outputs a full map for every joint. That matters for real-time feedback on a phone, which is why BlazePose uses heatmaps only while it’s being trained, then drops them and uses the faster direct regression when it actually runs.</p><h4>A short history</h4><p>Pose estimation went through several breakthroughs:</p><ul><li><strong>2014, DeepPose (Google):</strong> The first to use a deep convolutional neural network (CNN) for pose estimation. It used direct regression, then refined each joint in extra stages that zoomed in on it.</li><li><strong>2014 onward, heatmap models:</strong> Researchers found that predicting a heatmap for each joint worked better than direct regression, and most models that followed used heatmaps.</li><li><strong>2017, OpenPose (Carnegie Mellon University):</strong> Made real-time tracking of many people at once possible. It finds all the joints in the image first, then uses “Part Affinity Fields” to connect each joint to the right person.</li><li><strong>2019, HRNet:</strong> Improved accuracy by keeping a high-resolution version of the image throughout the network, instead of shrinking it and scaling it back up.</li><li><strong>2020, </strong><a href="https://www.research.google/blog/on-device-real-time-body-pose-tracking-with-mediapipe-blazepose/"><strong>BlazePose</strong></a><strong> (Google):</strong> Focused on speed. It gives up a little accuracy compared with the biggest models, but tracks one person’s 33 keypoints in real time on an ordinary phone, which is exactly what fitness apps need.</li><li><strong>2021, </strong><a href="https://blog.tensorflow.org/2021/05/next-generation-pose-detection-with-movenet-and-tensorflowjs.html"><strong>MoveNet</strong></a><strong> (Google):</strong> An ultra-fast model that detects 17 keypoints, with a “Lightning” version built for speed and a “Thunder” version built for accuracy. It was designed for fitness and movement apps running in a browser or on a phone.</li><li><strong>2022, </strong><a href="https://proceedings.neurips.cc/paper_files/paper/2022/hash/fbb10d319d44f8c3b4720873e4177c65-Abstract-Conference.html"><strong>ViTPose</strong></a><strong>:</strong> Showed that a plain vision transformer, built on the same transformer idea behind modern language models, works surprisingly well for pose estimation. Its largest version set a new accuracy record at the time.</li><li><strong>2023, </strong><a href="https://openmmlab.medium.com/rtmpose-the-all-in-one-real-time-pose-estimation-solution-for-application-and-research-6404f17cd52f"><strong>RTMPose</strong></a><strong> (OpenMMLab):</strong> Focused on real-world speed: its small version scored 72.2 AP (a standard accuracy score) on the COCO benchmark while running at over 70 FPS on a phone chip.</li><li><strong>2024, </strong><a href="https://github.com/facebookresearch/sapiens"><strong>Sapiens</strong></a><strong> (Meta):</strong> A foundation model trained on over 300 million images of people that tracks 308 keypoints (more on this in The future).</li></ul><p>Since 2020, the field has moved in two directions at once: bigger, more accurate models like ViTPose and Sapiens, and smaller, faster ones like MoveNet and RTMPose that run on phones. For a fitness app that has to give feedback in real time on the phone people already own, the second direction is the one that matters.</p><h3>From keypoints to rep counting</h3><p>Counting reps may seem complicated when all you have is joint positions, but a little trigonometry does most of the work.</p><h4>Measuring a joint angle</h4><p>To measure the angle at a joint, you use three keypoints. For the elbow, those are the shoulder (A), elbow (B) and wrist ©. You draw two vectors starting from the middle joint, B→A and B→C. The angle between them is the inverse cosine of their dot product divided by the product of their lengths:</p><p><strong>θ = arccos( (BA · BC) / (|BA| × |BC|) )</strong></p><h4>From angles to reps</h4><p>Once we have the angle, we can set <strong>thresholds</strong>: the angles a joint must reach for a movement to count. In a push-up, the elbow is close to straight at the top and around 90° at the bottom. A rep only counts when the angle goes through the full cycle, from the top, down past the bottom threshold, and back up. This logic is called a <strong>state machine</strong>, and it stops half-reps from being counted.</p><h4>Detecting movement vs. understanding good form</h4><p>With the joint positions and the angles between them, detecting most movements is possible. The harder problem is deciding whether a movement is done with good form, and what good form even is.</p><p>Take push-ups, one of the most popular exercises in calisthenics. Good form depends on several factors at once: keeping the whole body in a straight line, the angle between the arm and the torso, and the elbow angle (shoulder, elbow and wrist). Each of these has to be calculated, and each needs a target value that counts as correct. And that’s just one exercise. More complicated ones, like the pike push-up, add even more factors, so choosing these angles and the right values for them will be the hardest part of my project.</p><h3>My experiments</h3><h4>The setup</h4><p>I wanted to see for myself how reliable this technology is, so I ran a set of simple tests with Google’s MediaPipe Pose Landmarker on my laptop’s webcam, indoors. The program drew my skeleton, showed my elbow angles, marked each joint green (confident) or red (unsure), and displayed the frame rate (FPS).</p><h4>Results</h4><ul><li><strong>Holding my arm still:</strong> The angle jittered by about ±2°. That’s small, but a rep judge needs a margin of a few degrees around each threshold.</li><li><strong>Turning sideways:</strong> The arm farther from the camera turned red (unsure). Occlusion is real, so an app should ignore unsure joints.</li><li><strong>Hiding one arm or moving it out of frame:</strong> The model marked it as unsure. Confidence scores work as a safety signal.</li><li><strong>Distance from the camera:</strong> Every joint tracked well at 3 m, the farthest I tested. That’s enough space for most home exercises.</li><li><strong>Bright light behind me:</strong> Tracking still worked. A good sign, though I only tested indoors, not in direct sunlight.</li><li><strong>Fast movement:</strong> Tracking kept up, and on the full model, lag was barely noticeable. Real-time feedback is realistic.</li><li><strong>Lite vs. full model:</strong> Lite ran at 30–40 FPS and full at 25–30 FPS, but full gave steadier angles. Both are real-time; full is the better choice for judging reps on a laptop.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/710/1*_xDgR1nkycfZ4mvrJ3JFlQ.jpeg" /><figcaption>My visible arm (Elbow B, green) is tracked confidently at 63°. My other arm is out of frame, so the model marks its angle as unsure (red). Source: my own test with MediaPipe Pose Landmarker.</figcaption></figure><p>Overall, the technology is reliable enough to build on, as long as an app handles jitter and unsure joints carefully.</p><h4>What surprised me</h4><p>I honestly didn’t expect the model to work from 3 meters away, and tracking fast movement without any noticeable lag was genuinely impressive. That covers a lot of the situations a real user would be in, although I only tested it indoors and on myself.</p><h3>Limitations</h3><h4>Occlusion</h4><p>When a body part is hidden by the body itself, by furniture, or by being out of frame, the model has to guess where it is. Google lists this as one of the main challenges for fitness use (<a href="https://www.research.google/blog/on-device-real-time-body-pose-tracking-with-mediapipe-blazepose/">source</a>), and my sideways test showed it clearly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/520/1*O-PCwe2LB_0j_gRkAznQMA.jpeg" /><figcaption>Turned sideways, my far arm is hidden behind my body, so its elbow and wrist turn red: the model is guessing where they are. Source: my own test with MediaPipe Pose Landmarker.</figcaption></figure><h4>Small body parts</h4><p>Hands and feet take up little of the image and are harder to track. A <a href="https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10593217/">study</a> that used BlazePose to measure joint angles had to leave out hand and foot movements because they couldn’t be tracked consistently.</p><h4>No real depth</h4><p>Most phone models work in 2D, so they can’t truly see depth. The same elbow bend can measure differently depending on where the camera is, which is why a side view works better for some exercises.</p><h4>Speed vs. accuracy</h4><p>Bigger, more accurate models are slower. I measured this myself: the full model gave steadier angles but ran 10–15 FPS slower than the lite model.</p><h4>Fairness</h4><p>It’s hard to know whether a model works equally well for everyone. As <a href="https://ai.sony/blog/Exposing-Limitations-in-Fairness-Evaluations-Human-Pose-Estimation">Sony AI</a> points out, many datasets don’t include labels such as age, body type or disability, so teams often can’t even test whether their model is biased. My hypothesis is that people who are less represented in training data may be tracked less accurately, but this needs to be tested.</p><h4>Who decides a “good rep”?</h4><p>The rules that judge a rep are written by people. If the thresholds are too strict, the app fails good reps; too loose, and it passes bad ones. The only real way to know is to check the app’s judgments against a qualified trainer’s.</p><h4>Privacy</h4><p>Many people don’t want to be filmed while exercising. The best solution is to run the model on the device itself and store no video at all, keeping only the joint positions and results.</p><h3>Who’s working on this</h3><p>Big tech companies and research labs build the models, and a growing number of apps build on top of them.</p><ul><li><strong>Google:</strong> Builds <a href="https://www.research.google/blog/on-device-real-time-body-pose-tracking-with-mediapipe-blazepose/">MediaPipe and BlazePose</a>, the fast on-device models many fitness apps use.</li><li><strong>Meta:</strong> Built <a href="https://github.com/facebookresearch/sapiens">Sapiens</a>, a large foundation model for human vision.</li><li><strong>Stanford University:</strong> Built <a href="https://engineering.stanford.edu/news/opencap-sophisticated-human-biomechanics-smartphone-video">OpenCap</a>, which measures movement and forces from smartphone video.</li><li><a href="https://apps.apple.com/app/id6748878622"><strong>MyRepsCount</strong></a><strong>:</strong> Counts reps in real time for 30+ exercises using the phone’s camera, and says it never records, stores or sends the camera feed.</li><li><a href="https://apps.apple.com/app/id6756504196"><strong>RepSquad (AI Rep Counter)</strong></a><strong>:</strong> Runs MediaPipe on the device to count reps and gives an A/B/C form grade.</li><li><a href="https://apps.apple.com/us/app/id1555380908"><strong>ReadyFit</strong></a><strong>:</strong> Lets users take official US military fitness tests by recording videos, which are then verified and scored by human graders.</li></ul><h3>The gap</h3><p>After looking at these apps, I noticed four gaps:</p><ol><li><strong>Counting and grading, not explaining.</strong> From the descriptions, most apps focus on counting reps or giving an overall grade. Few clearly promise to explain, rep by rep, what went wrong and how to fix it, which is what a beginner actually needs.</li><li><strong>Official standards.</strong> ReadyFit judges official fitness tests, but with human graders. The camera-based apps use AI, but not against official standards.</li><li><strong>Proof.</strong> None of the app descriptions I read say how often their judgments agree with a qualified trainer.</li><li><strong>Local data.</strong> I couldn’t find UAE-specific data on training injuries, so the size of the problem here is still unknown.</li></ol><h4>Where my app fits</h4><p>Gaps 1 and 3 are the ones my app will focus on. The main reason I’m building this is to help beginners improve their form, and counting or grading alone won’t do that. Beginners need to know exactly what’s going wrong, and that’s what I want to provide. To make sure the app is actually reliable, its judgments need to be checked against a qualified trainer’s.</p><h3>The future</h3><h4>From joints to real biomechanics</h4><p>Stanford’s <a href="https://engineering.stanford.edu/news/opencap-sophisticated-human-biomechanics-smartphone-video">OpenCap</a> uses video from two smartphones to measure not just how a person moves, but also the forces in their muscles and joints. It takes minutes instead of days, at less than 1% of the cost of a traditional $150,000 motion lab. In one study, it <a href="https://news.stanford.edu/stories/2025/11/smartphones-monitoring-neuromuscular-disorder-treatment-research">distinguished between neuromuscular diseases</a> with 82% accuracy, compared with 50% for traditional stopwatch-timed tests, and Germany’s national volleyball team used it to evaluate sports injuries in 160 athletes.</p><h4>Foundation models for the human body</h4><p>Just as large language models learned from huge amounts of text, Meta’s <a href="https://github.com/facebookresearch/sapiens">Sapiens</a> was pretrained on over 300 million images of people. It detects 308 keypoints, with detailed points on the face and hands, and its performance kept improving as the model grew from 0.3 to 2 billion parameters.</p><h4>Big models vs. small devices</h4><p>The most accurate systems need either more cameras (OpenCap uses two) or a lot of computing power (Sapiens goes up to 2 billion parameters), which a phone can’t easily provide. Phone models like BlazePose are fast but less accurate. As models become more efficient, more of this power will likely run directly on phones.</p><h4>Using motion over time</h4><p>Newer models look at sequences of frames instead of processing each frame on its own. This can reduce jitter and help the model follow body parts through short moments of occlusion.</p><h4>What this means for my project</h4><p>Each of these makes something possible that wasn’t before. OpenCap shows that phones can go beyond angles and estimate the forces on joints, which could one day let an app warn you when a movement puts too much strain on your shoulders or knees, instead of just counting reps. Sapiens tracks far more detail on the hands, one of the body parts I listed as a limitation; for calisthenics skills like handstands, where hand and wrist position matter, that would be a big step. As models become more efficient, that kind of accuracy could run on a phone, so my app could get more accurate without giving up on-device privacy. And models that use motion over time could reduce the jitter and occlusion problems I saw in my own tests, which means fewer reps judged wrongly.</p><h3>My hot take</h3><p>My hot take is that pose estimation won’t replace knowledgeable coaches, but within the next few years it will give beginners who have no coach at all consistent feedback on their form. I believe this for two reasons. First, the hardest technical part is already solved: in my own tests, the model tracked my joints steadily, to within about ±2°, in real time on an ordinary laptop. Second, beginners already own the only equipment it needs, a phone camera, while a personal coach costs money many beginners can’t or won’t spend. What’s still missing isn’t better tracking but trust: proof that an app’s judgments match a real coach’s. The apps that prove that will be the ones beginners rely on.</p><h3>Conclusion</h3><p>Pose estimation has moved from research labs to something that runs on an ordinary phone. The AI can already find the body’s joints in real time; the next challenge is turning those joints into feedback that beginners can trust.</p><p>So, back to the question from the beginning: can pose estimation be trusted to coach beginners? My research and tests show it’s reliable enough to track the body and give basic form corrections. Whether it can judge good form as well as a trainer is exactly what my project will test.</p><h3>What’s next</h3><p>In the replicate phase of my project, I will build my own version from scratch using Google’s MediaPipe model. It will track the joints, calculate the angles, count reps for multiple exercises, and flag basic form mistakes.</p><p>In the application phase, I will build on that in three ways. First, I’ll focus on calisthenics skills that are harder to judge than basic exercises, like pike push-ups, pull-ups, dips and handstand holds. Second, instead of fixed messages, an LLM will explain why each rep failed and how to fix it, in Arabic or English. Third, I’ll check my app’s judgments against a qualified coach to see how often we agree, so beginners know whether they can actually trust it.</p><h3>About me</h3><p>I’m Mohammed Ebrahim AlZaabi, a 12th-grade student in Ras Al Khaimah and a member of The Knowledge Society (TKS). My main focus area is artificial intelligence, and I’m interested in using AI to make people’s lives easier. With this project, I hope to make training more accessible and safer for beginners.</p><h3>References</h3><ol><li><a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC11297712/">Exercise-related injuries among gym members in Saudi Arabia (2024)</a></li><li><a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC5315259/">Injuries in an extreme conditioning program</a></li><li><a href="https://doi.org/10.1145/3603618">Deep Learning-based Human Pose Estimation: A Survey (ACM Computing Surveys, 2023)</a></li><li><a href="https://www.research.google/blog/on-device-real-time-body-pose-tracking-with-mediapipe-blazepose/">On-device, real-time body pose tracking with MediaPipe BlazePose (Google Research)</a></li><li><a href="https://ai.google.dev/edge/mediapipe/solutions/vision/pose_landmarker">Pose Landmarker guide (Google MediaPipe documentation)</a></li><li><a href="https://blog.tensorflow.org/2021/05/next-generation-pose-detection-with-movenet-and-tensorflowjs.html">Next-generation pose detection with MoveNet (TensorFlow Blog, 2021)</a></li><li><a href="https://proceedings.neurips.cc/paper_files/paper/2022/hash/fbb10d319d44f8c3b4720873e4177c65-Abstract-Conference.html">ViTPose: Simple Vision Transformer Baselines for Human Pose Estimation (NeurIPS 2022)</a></li><li><a href="https://openmmlab.medium.com/rtmpose-the-all-in-one-real-time-pose-estimation-solution-for-application-and-research-6404f17cd52f">RTMPose: real-time pose estimation (OpenMMLab, 2023)</a></li><li><a href="https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10593217/">A webcam-based machine learning approach for three-dimensional range of motion evaluation (PLOS ONE)</a></li><li><a href="https://ai.sony/blog/Exposing-Limitations-in-Fairness-Evaluations-Human-Pose-Estimation">Exposing limitations in fairness evaluations: human pose estimation (Sony AI)</a></li><li><a href="https://engineering.stanford.edu/news/opencap-sophisticated-human-biomechanics-smartphone-video">OpenCap: human biomechanics from smartphone video (Stanford Engineering)</a></li><li><a href="https://news.stanford.edu/stories/2025/11/smartphones-monitoring-neuromuscular-disorder-treatment-research">Smartphones democratize neuromuscular disease monitoring (Stanford Report)</a></li><li><a href="https://github.com/facebookresearch/sapiens">Sapiens: Foundation for Human Vision Models (Meta, ECCV 2024)</a></li><li><a href="https://apps.apple.com/app/id6748878622">MyRepsCount on the App Store</a></li><li><a href="https://apps.apple.com/app/id6756504196">RepSquad (AI Rep Counter) on the App Store</a></li><li><a href="https://apps.apple.com/us/app/id1555380908">ReadyFit on the App Store</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=325d17dc24f6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Building Maps to a Million Plays]]></title>
            <link>https://medium.com/@m.ebrahim.alzaabi/what-i-learned-making-fortnite-games-played-over-a-million-times-e7d817712b9a?source=rss-7ad183266eb7------2</link>
            <guid isPermaLink="false">https://medium.com/p/e7d817712b9a</guid>
            <dc:creator><![CDATA[Mohammed Ebrahim Alzaabi]]></dc:creator>
            <pubDate>Tue, 29 Sep 2026 16:42:16 GMT</pubDate>
            <atom:updated>2026-09-29T16:48:55.129Z</atom:updated>
            <content:encoded><![CDATA[<p><em>How I went from making maps with in-game devices to running games played in 128 countries.</em></p><p>About five years ago, I started making Fortnite maps. There was no big plan behind it. I liked the game, I liked building things, and Fortnite Creative let me do both.</p><h3>Starting with devices</h3><p>Back then, Unreal Editor for Fortnite (UEFN) didn’t exist yet. Everything in Fortnite Creative was built with devices, the in-game tools that control what happens in a map, with no code at all. I learned by building, and I built a lot. Every map taught me something about what made players stay and what made them leave.</p><p>The first real breakthrough came when two of my maps passed Epic Games’ review for featured placement in Fortnite Discover: مدرسة المشاغبين (School of the Troublemakers) and سباق الخطاف بين العوالم (Grappling Hook Race Between Worlds).</p><h3>A new tool, a new team</h3><p>When UEFN arrived, maps could use Verse, Epic’s programming language, and suddenly much bigger games were possible. I teamed up with a friend who wrote most of the Verse, while I led the project: building the maps, setting up the devices, planning updates and tracking how players behaved.</p><p>Our first big release was Space Ship Tycoon, in April 2024. It took off immediately, hitting 74,400 plays in a single day during its first week and going on to 206,000 plays in total. Watching those numbers climb was one of the best feelings I’d had making games, and it made me want to aim even higher with the next one.</p><h3>Supermarket Simulator Tycoon</h3><p>That next one was Supermarket Simulator Tycoon, released on 12 September 2024, and it was the most ambitious thing I’d built.</p><p>My earlier maps were mostly for Arabic-speaking players. This time I built in English and localised the game into every language Epic supports, because I wanted it to be for everyone. It worked: players came from 128 countries.</p><p>We also kept the game moving. Over the next five months, we shipped 22 releases, about one a week, and players kept coming back for them. More than half of all plays came after the first month. The game went on to reach 88.6 million impressions, a best day of 32,500 plays, and eventually crossed a million plays 1.04 million in total.</p><h3>Where it stands now</h3><p>Looking back, it’s strange to think it all started with devices and an idea. Both games are still live if you want to try them: Supermarket Simulator Tycoon (3464–1480–4704) and Space Ship Tycoon (7267–6425–7527).</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e7d817712b9a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>