Baseball as an engineering problem

Pitching Biomechanics

A baseball problem turned into a wearable-sensor and machine-learning project for studying pitching motion.

Pitching made the question personal: could low-cost sensors capture enough movement to make useful feedback more accessible? I connected IMUs to an Arduino, collected wireless throw data, and trained a sequence model to study how motion changes across a pitch.

Hardware: Arduino Uno R4 WiFi + 2 MPU6050 IMUs
Dataset: 20 players, 400+ throws
Model: BiLSTM sequence model
Focus: Pitch velocity + movement patterns

Christopher wearing wired motion sensors while holding a baseball during throwing-data collection. — The collection setup was intentionally simple: wearable sensors, a ball, and repeated throws.

Two time-series charts showing hip IMU acceleration and angular velocity during a throw. — Each throw became a time-series signal with a visible burst around the pitching motion.

Heatmap showing attention-weighted sensor channels across time for a baseball pitch. — The model made parts of the motion sequence easier to inspect, not just predict.

System flow

IMU sensors
Arduino Uno R4 WiFi
Wireless transfer
Python CSV collection
TensorFlow pipeline

Bidirectional LSTM diagram showing forward and backward sequence layers through time. — Sequence models are useful here because throwing is a time-based motion, not a single static pose.

Data collection video

Pitching Biomechanics data collection video

Shows the throwing-data collection setup behind the wearable-sensor project.

Open on YouTube

Behind the scenes video

Pitching Biomechanics behind-the-scenes video

Shows additional project context from the build and testing process.

Open on YouTube

Problem

Why throwing?

Throwing is a coordinated full-body movement involving the legs, hips, torso, and arm. Most players cannot casually access expensive biomechanics tools, so meaningful feedback often comes down to what a coach can see in the moment.

I approached that as a measurement problem: wearable IMU sensors capture the movement, embedded firmware moves the data wirelessly, and machine learning helps make the motion sequence easier to study after the throw.

Model

What the model learned

The model treated each pitch as a time sequence rather than a single snapshot. That mattered because the useful signal lives in timing: when the body accelerates, how the motion changes, and where the most informative moments appear.

I kept the exact model comparisons in the linked poster. The important portfolio takeaway is the workflow: collect motion, preserve the sequence, and use the model to inspect patterns connected to velocity and mechanics.

Applications

Where this could go

The same basic approach could apply beyond baseball wherever motion is hard to observe directly: rehabilitation, ergonomics, gait analysis, or other sports movement.

This implementation stayed anchored in pitching. That boundary is important: it made the project specific enough to test honestly while pointing toward a broader idea for accessible motion feedback.