Gait Event (GE) detection (e.g. initial contact and final contact) is a fundamental component
of gait analysis, enabling the computation of clinically relevant spatiotemporal parameters (SGPs) such
as stride time, stance phase, and gait variability. These metrics are widely used in the assessment of
neurological disorders, rehabilitation monitoring, and fall risk estimation. Traditionally, GEs are
identified using force plates or motion capture systems (MoCaps), which provide high accuracy but
are restricted to controlled lab environments.
In recent years, inertial measurement units (IMUs) have emerged as a portable and cost-effective
alternative for real-world gait monitoring. While many studies rely on multiple sensors (e.g. bilateral
foot or shank), reducing the system to a single IMU on the sacrum is particularly attractive due to its
ease of use and patient compliance for remote monitoring. However, this configuration introduces
additional challenges, as the sacrum signal is less directly linked to discrete GEs.
Deep learning (DL) approaches have shown strong potential for IMU-based GE detection by learning
features directly from raw time-series data. Prior work [1][2][3] has demonstrated promising results
using single IMUs placed on the shanks or sacrum, but several limitations remain: performance is often
not analyzed across different gait populations, GEs may be inferred indirectly from predicted gait
phases, and existing single-sensor approaches do not systematically investigate how input signal
representations affect detection performance.
As a result, the application of Temporal Convolutional Networks (TCNs) for GE detection from a single
sacrum-mounted IMU, and the role of different input signal configurations in this setting, remain
insufficiently explored.

To develop and evaluate a TCN for GE detection using single sacrum-mounted IMU signals
from an in-house dataset. The study will specifically investigate how different input signal
configurations affect model performance through an ablation study, including:
• Vertical acceleration only
• 3D acceleration
• Combined acceleration and gyroscope signals
The goal is to determine if TCNs can be extended to sacrum-mounted IMUs for GE detection and which
inputs allow the model to learn the optimal information.

Literature Review (ETOC: 2 months): Review existing methods for gait event detection using
IMUs, with a focus on DL approaches (TCNs) and sensor placements (shank, feet,
sacrum)
- Dataset Familiarization (ETOC: 1 month) - Understand the structure and characteristics of the
IMU dataset to be used.
- Implementation of TCN Pipelines (ETOC: 4 months): Design and implement a TCN framework
tailored for GE detection from sacrum IMU data. Adapt architectures from existing work [1][2]
to the new sensor placement.
- Ablation Study and Evaluation (ETOC: 2 months): Systematically compare different input
configurations, including vertical acceleration only, 3D acceleration, and combined
acceleration and gyroscope signals. Evaluate performance in terms of GE detection accuracy
and timing error, and analyze which input representation is most suitable for a single sacrummounted
IMU.

• Following a MSc in a field related to one or more of the following: Computer Science, Biomedical
Engineering, Applied Computer Science - Digital Health.
• Strong programming skills (Python).
• Ability to write scientific reports and communicate research results at conferences in
English.