A rising number of people in our community, especially elderly people, suffer from gait disorders and gait deficits. Most clinics and hospitals which have the prerequisite of a very high daily patient throughput typically rely on force plates and cost-effective two-dimensional gait analysis tools to determine ground reaction forces (GRF) and gait kinematics during locomotion to quantify the patients' gait disorders and to evaluate the patients' progress during physical therapy treatment. From the shapes of the GRF signals and from visual observation the expert tries to diagnose whether a functional deficit exists and which type of deficit the patient suffers from.
The manual analysis of gait patterns by visual observation and by visual inspection of GRF signals, however, leads to subjective and non-repeatable assessments. Furthermore, it has been shown that several parameters (e.g. walking speed, age, gender) influence the shape of gait patterns significantly and further impede the manual (visual) assessment and comparability of different patterns. Thus, more objective means to support the therapist in the assessment of gait patterns are required. Automatic analysis methods enable novel ways to obtain objective and repeatable assessments and thus represent a valuable supporting tool for the expert in making diagnoses.
Recently, there has been an increase of research efforts on the automatic analysis and classification of human gait. From related work, however, we observe that the performed analyses and evaluations are often applied on rather small (and artificial) datasets and that the developed methods are usually tuned manually to one particular or a few deficits only. Thus, the general applicability of existing approaches in a real-world scenario is limited.
In the proposed project, we are provided with a large-scale database of real-world gait patterns captured during clinical praxis. This dataset enables us to develop novel methods for automated gait analysis that better take the real-world requirements into account.
The contributions of the proposed project are:
- The development of robust normal behavior models that are invariant to different walking speeds, ages, genders, and physical conditions of the patients to enable the distinction of normal from abnormal gait;
- The introduction of novel features for gait data that are learned directly from the data. The novel features allow for the classification of a wide range of functional deficits and should support the expert in knowledge discovery;
- Data-driven access to gait pattern databases (by similarity retrievaltechniques) enables immediate access to a much larger data basis than currently possible.
The vision beyond this project is to develop methods which are capable of effectively supporting medical specialists in making decisions, increasing patient throughput and in consequence reducing medical treatment costs.