Skip to main content

Project Suggestions

Concrete project pitches presented in CACoM lectures by instructors or external speakers. Each is a starting point — narrow enough to act on, but with deliberate room for your team to shape the question, the methods, and the evaluation. If none of these fit, see the broader Areas of Interest; once you've chosen something, work through the Refinement Guide to turn it into a concrete proposal.

note

This list grows as new pitches are given. If you remember a topic from a lecture that isn't here yet, ask the instructors and we'll add it.

🦶 Foot Sensitivity in Barefoot Shoes​

"Barefoot" or minimalist shoes use very thin, flexible soles, intended to preserve sensory feedback from the ground — in contrast to thick-soled trainers, which cushion the foot but may also alter how we walk over time. A previous CACoM team built a thin-leather Fullsoul barefoot shoe with an embedded foot-sole pressure sensor and compared its pressure distribution against a sneaker carrying a similar sensor inside its thicker midsole. The motivating hypothesis is that thick-soled footwear changes how we walk — and not necessarily for the better.

This project goes one step further and asks: how much sensory feedback does the foot actually receive in each condition? A quantitative answer would let us compare barefoot, Fullsoul, and competing minimalist designs on the property they all claim to preserve.

Conditions to compare:

  • Bare foot.
  • Foot inside the Fullsoul (thin leather sole).
  • Foot inside the Fullsoul with the velostat pressure sensor in place — to confirm the sensor itself does not blunt sensitivity.
  • Foot inside a thin-rubber-soled barefoot shoe (the most common alternative on the market).

Established method — tuning fork (courtesy of Prof. Dr. Walther, Schön Klinik): a tuning fork is set vibrating and pressed against the foot; the subject marks on a foot schematic where the vibration was felt. Reproducible, low-cost, and clinically grounded.

Find or develop further sensitivity methods. A single test rarely captures the whole picture, and a quantitative comparison improves with multiple independent measures. As one starting point, Daniel Lieberman's barefoot-locomotion work has used tactile sensitivity tests on subjects lying flat on the floor — but lying down may itself reduce sensitivity, so any adaptation needs to think carefully about posture.

Additional tasks:

  • Further benchmarking of the existing pressure sensor.
  • Adapt the embedded code and sensor link to use Bluetooth Low Energy (currently wired), so the shoe can be worn freely outside the lab.

❤️ Heartwise — Wearable-Based Early Detection of Heart Insufficiency​

Heartwise is a funded study starting soon in Singapore that aims to build a wearable-based early-warning system for recurrence of heart insufficiency. The premise is simple: cardiac events tend to recur in patients who have already had one, and a pipeline that picks up early signals could close the gap to clinical intervention.

Enrolled patients will wear an actibelt® effectively 24/7 (swapping it once a day for a freshly charged unit), record a daily weight measurement, and capture a short set of body sounds — primarily heart and lungs — with an electronic stethoscope. Together, these streams are the input that the long-term Heartwise model is meant to learn from.

The CACoM project is an early pilot for that larger study, in two stages.

Stage 1 — Self-experimentation (3–7 days)​

Each team member runs the same protocol on themselves for several days: actibelt all day, daily weight, daily stethoscope recordings. There is no expectation of mocking heart insufficiency or any other pathology. Instead, the aim is curiosity-driven exploration of what the data even contain. Useful questions to push on:

  • What do I actually hear through the stethoscope? Can I isolate the heartbeat in the signal?
  • Can I separate heart sounds from lung sounds — or pick up blood flow at all? Any audible murmurs?
  • Can I couple actibelt activity with heart rate or sound character? After an intense workout, a meal, and an hour of rest, is my heart rate still elevated? Do my heart sounds change?
  • What concise summaries or visualizations capture the day-to-day variation cleanly?

Stage 1 has two purposes. It teaches the realities of working with these signals — sensor placement, noise, alignment, file formats — in a forgiving setting. And it forces the processing and visualization decisions that would otherwise have to be made under time pressure once real patients are involved.

Stage 2 — Real patients at TUM Klinikum Rechts der Isar​

Through the telemonitoring centre at TUM Klinikum Rechts der Isar, the same protocol can be run on patients with diagnosed heart insufficiency. This is an unusual opportunity for a course project — direct contact with real clinical data, in a clinical setting, on a real clinical question, and with the Heartwise study as the eventual home for whatever pipeline you build in the process.

caution

Stage 2 involves real patient data and adds significant coordination overhead. Review Data Privacy & Responsible Data Handling and the Collaboration Policy before designing your protocol, and discuss the timeline with instructors early.

⏱️ Cross-Device Clock Synchronization​

Any recording device with a free-running clock drifts. The longer the recording, the worse the drift. The actibelt® is calibrated at the start of each session, but since it can record continuously for up to two months, accumulated drift becomes a real obstacle to any multi-device analysis — for example, comparing two actibelts on the same subject, or fusing actibelt data with a phone's IMU stream.

Manual re-synchronization is impractical at scale. The goal of this project is to build an automatic cross-device synchronization method, generalizable beyond actibelts to any recording device that lacks a central time source.

Approach:

  • Mount two devices rigidly together — a 3D-printed jig is the cleanest way to lock them into a shared frame of reference — and execute a deliberate motion pattern (rotations, taps, a known sequence).
  • Develop an algorithm that scans both devices' three-axis acceleration streams, identifies the shared motion signature, and aligns the two time bases from it.
  • Two-actibelt synchronization is the natural starting point; actibelt + smartphone is the more useful long-term target, since it lets the phone act as a wider-purpose reference.

The success criterion is clean and quantifiable: how well do the aligned signals agree post-sync, and how stable is the alignment as recording length grows? Squarely a signal-processing and time-series-alignment problem (cross-correlation, dynamic time warping, and friends).

🎥 Automated Gold-Standard Annotation for Walking Data​

Training and validating advanced movement models requires gold-standard walking data — recordings annotated with the exact timing of every heel strike and toe-off, plus a reliable speed reference. Today, producing one such recording takes four people:

  1. The subject, walking while wearing an actibelt.
  2. A second person walking alongside with a measuring wheel (also carrying an actibelt) to provide ground-truth speed.
  3. A third person filming the subject's steps with a camera.
  4. A fourth person who, after the fact, scrubs through the video frame by frame and records the exact timing of each heel strike and toe-off.

Person 4 is the bottleneck. Eliminating that role would dramatically speed up data collection and make training of better movement models tractable.

The project asks you to replace person 4 with an automated annotation pipeline built on off-the-shelf computer vision. Two complementary approaches are worth considering:

  • Pose estimation + step logic. Use a standard pose estimator — e.g. via the ultralytics Python package — to recover a stick-figure trajectory of the subject, then write algorithms that detect heel strikes and toe-offs from the joint kinematics over time.
  • Direct event detection. Train or fine-tune a model to predict heel-strike and toe-off frames directly from the video, sidestepping the pose-estimation intermediate.

A good outcome should extend naturally toward the longer-term vision: automatic activity labelling (walking, sitting, lying, transitions) from the same video stream, opening the door to gold-standard annotation in less constrained, free-living datasets.