Animal Motion Event Detection Study

About the project We are running an animal behavior study with a custom wearable motion logger. An observer marks specific behavioral events in real time. We want to find out whether these events can be detected automatically from the motion signals and, if so, build a detector that runs in real time. Further background will be shared after an NDA is signed. The key question Can event moments be reliably distinguished from non-event moments at a similar level of physical activity, using tail and body motion alone? Data we provide - Sensors: two MPU6050 IMUs (3-axis accelerometer + 3-axis gyroscope), one on the tail and one on the torso as a body-motion reference. Both are sampled in the same tick (<1 ms skew). - Settings: 100 Hz sampling, on-chip low-pass filter at 20 Hz, accelerometer ±8 g, gyroscope ±2000 dps. The actual scale is written in each file header (some early pilot files used ±1000 dps). - Format: one CSV per 3-minute segment with raw int16 values. Each file header contains the session's gravity vectors and gyro bias from a 2-second stillness calibration, plus a calibration quality flag. - Quality flags: a per-row status bitmask marks sensor read failures and timing gaps. - Labels: an experienced observer presses a wireless button at the moment the animal first shows the response of interest, not at a later confirmation behavior. Radio latency is compensated, so the remaining offset is the observer's reaction time, typically under 1 s. A second button marks the start and end of each trial, so non-event windows can be taken from known clean segments. - Control runs: we also record identical trials in which the stimulus is absent, so no event can occur. - Amount: pilot recordings from one animal are available now. Data collection is ongoing, and new recordings will be delivered in batches. More animals may be added later. Scope of work Phase 0 – Data quality review (approx. 3–5 days) Using 2–3 sample sessions: - Check noise level, dropped samples, saturation and sensor mounting artifacts - Check how consistently the signals change around the labeled events - Recommend changes to sensor settings, mounting or data collection if needed Deliverable: short report with plots and recommendations. Phase 1 – Feasibility (approx. 2–4 weeks, depending on data availability) - Preprocessing: unit conversion, bias removal, gravity/orientation alignment using the header calibration, filtering, windowing - Feature extraction, for example: tail motion frequency (main band expected around 1–6 Hz) and amplitude, left-right asymmetry, tail elevation, stillness periods, and body-sensor features to separate tail motion from whole-body movement. Other approaches (e.g. deep learning on raw signals) are welcome if justified. - Train and evaluate a classifier using leave-one-session-out validation (sessions recorded on different days), since data currently comes from a single animal - Report sensitivity, specificity, false alarms per hour and detection latency, using a tolerance window around each label Deliverable: Python code, documented notebook, and a technical report with a clear answer to the key question. A well-supported negative result is an acceptable outcome. Phase 2 – Real-time implementation (only if Phase 1 is positive; scope and price agreed separately) A real-time version of the detector suitable for ESP32 (or an Android phone), with code delivery and integration support. How we will work - Communication in English, with a short weekly update (message or video call) - Data shared via a secure link after the NDA is signed - Payment per milestone, released on acceptance of each phase's deliverable Required skills - IMU / inertial signal processing - Activity or behavior recognition from wearable sensors - Time-series classification / machine learning in Python - Nice to have: embedded (ESP32) experience, animal data experience Terms - NDA and full transfer of IP rights for all code and outputs are required. - Data and further details will be shared only after the NDA is signed. To apply, please include 1. Examples of similar work (publications, GitHub, etc.) 2. A half-page outline of your approach to Phase 0 and Phase 1 3. Separate time and cost estimates for Phase 0 and Phase 1