Fitness technology has come a long way over the last few decades, taking advantage of the latest technologies, such as wearables to keep tabs on user’s health and biofunctions and algorithms that monitor and track physical activities. AI is also being utilized to help augment exercise routines, enhance yoga positions and prevent injuries while working out. Hackster’s February Impact Spotlights series webinar on Fitness Technology highlights a trio of talented individuals and how they took advantage of AI to augment their fitness projects.

First guest Whitney Knitter described how she took advantage of AI to judge her poses while practicing yoga. If not practiced correctly, the wrong poses can affect posture, which can lead to sore muscles and other adverse effects. To help her achieve the correct pose posture, Knitter developed an AI-based pose estimation platform to analyze her stances using pre-trained AI models that compare her poses with the correct ones.

Knitter designed her AI-based platform using a Texas Instruments SK-TDA4VM Edge AI starter kit and Edge AI SDK, which provided a pre-trained 6D pose estimation model based on YOLOX from TI’s Model Zoo repository. The model infers poses directly from images without any complex postprocessing, and TI’s Tidal tools optimize them for hardware acceleration using the SK-TDA4VM. Knitter customized the model using TI’s Model Analyzer to compile it with calibration images of her own yoga poses. She then fine-tuned the system by ensuring image sizes matched expectations and corrected inference-related errors, such as misaligned skeleton overlays. Knitter then optimized execution by correctly using TIDLCompilationProvider for calibration and TIDLExecutionProvider for inference.

Knitter’s yoga platform successfully identified key points in her poses, printed slope values of skeleton lines for each body part and correlated them to specific poses for real-time yoga feedback.

Justin Lutz was on hand to provide insight into his Arduin-Row device, which acts as a rowing machine coach that provides feedback using the rower’s tempo and how the rowing handle moves through the stroke. Lutz designed the Arduin-Row using an Arduino Nicla Sense ME to garner accelerometer data and Edge Impulse to train a TinyML model to classify rowing stroke types and detect anomalies in handle placement.

Lutz collected 18 minutes of data to train the TinyML model of three rowing states – easy, low strokes-per-minute (SPM) and high SPM, and the anomaly detection was used to determine if the rowing handle remained level through each stroke. The trained model was then deployed as an Arduino library and integrated into the Arduino IoT Cloud to provide real-time feedback and remote monitoring of Lutz’s performance. He then used the “Coach’s Oder” chat-based interface to provide the rowing feedback.

Lutz also took the opportunity to monitor CO2 levels during his rowing workouts using an eCO2 sensor to check local air quality, which is important when in confined places with no airflow.

Avnet’s Mario Bergeron rounded up the webinar with the mechanics of rock climbing using AI. As avid rock climbers, Bergeron and his wife wanted a way to maximize their rock-climbing abilities, taking inspiration from rock climbing as a relatively new sport at the 2021 Olympics. To accomplish that goal, Bergeron turned to AI to find the ideal pose to make climbing more efficient and gain a competitive edge.

Early AI models weren’t up to the task of detecting the complex poses climbers use to navigate rock faces due to a lack of training data. To overcome that challenge, Bergeron took advantage of YOLOV7 – an accurate real-time object detection platform based on convolutional neural networks. The primary objective of his climbing project was to automate side-by-side comparisons of climbing performances. The process of developing that capability has been tedious; however, Bergeron found promise in combining machine learning and computer vision.

Bergeron’s climbing project features two main components – Video Annotation, which applies pose estimation to annotate climbing footage, which was done via pyOpenAnnotate, and Side-by-Side Viewing, which aligns and synchronizes climbing videos. For data collection, Bergeron recorded 48 climbing videos captured using an iPhone 14 Pro in accelerated mode to save battery life; however the Deep SORT tracking algorithm he used failed to consistently track the climbers due to using that accelerated mode (frame skipping), sudden changes in the climber’s position and misidentification of rock features.

Bergeron plans on testing the platform on videos with 30 FPS to see if tracking improves and implement a background subtraction technique to focus tracking on the climber rather than the environment.