Home | english  | Impressum | Datenschutz | Sitemap | KIT

Deep-learning supported gait analysis

Deep-learning supported gait analysis

Anees Qumar Abbasi,Werner Nahm

Gait analysis is an important aspect for the clinical assessment of certain neurological diseases, as e.g. Hemiparesis, Parkinson or Multiple Sclerosis (MS). The pathology of these diseases is usually manifested by a certain gait disorder. The measurement of predefined gait parameters can give an estimation about the progress of the disease.
The human gait can be classified into several gait phases. The most important ones are the initial contact, the loading response and the terminal stance. As can be seen from the figure below, there are certain ankle, knee and hip angles which are considered to characterize a normal gait pattern.
Nowadays, patients with gait anomalies need to visit a gait laboratory from time to time in order to get their gait assessed with the help of special software tools (usually marker-based) and human experts (doctors and physiotherapists) which are present in the laboratory only. However, the assessment of relevant gait parameters usually takes places manually by analyzing the video streams. 
Additionally, patients usually learn techniques in order to maintain the best possible gait concerning their condition. Still there is no online monitoring of the patients gait in the real-world environment so it is hard to know whether the patient really implements the techniques he learns in the laboratory.
In the scope of this project, we therefore aim for a marker-free video-based gait analysis under real-world conditions. The idea is to provide a system, which can automatically assess human gait parameters based on self-made smartphone videos. The system should be capable of detecting the relevant gait phases and calculating the crucial gait parameters out of it. 
For achieving the goals of the project, there is a need to make research within the field of the latest image and video processing algorithms. We especially aspire to make use of Machine Learning/Deep Learning methods in order to extract the relevant features from the videos. The latter have shown to beat more classical approaches in the relevant subareas of time series analysis and classification, object detection and pose estimation.