A 3D Deep Learning Approach for Classification of Gait Abnormalities Using Microsoft Kinect V2 Sensor

In this paper, a deep learning approach is proposed based on a 3D Convolutional Neural Network for the classification of gait abnormalities. Six gait classes are considered, including Trendelenburg, Steppage, Stiff-legged, Lurching, and Antalgic gait abnormalities as well as normal gait. The proposed scheme is applied to a recently-published dataset from the literature. This dataset consists of the gait data recorded by multiple Microsoft Kinect v2 sensor from 25 joints of a person during walking on a specified walkway. In this dataset, for each of the 6 gait classes, ten people have attended the data collection procedure; and for each participant, 120 walking instances have been recorded. Each instance includes the spatial and temporal information of the walking, and it is converted to two 3D images, which respectively display the changes of the Coronal (X-Z) and Sagittal (Y-Z) views of the originally captured data over time. These two 3D images are used as the input of the proposed 3D convolutional neural network. There are a total of 14400 3D images in this dataset. In order to demonstrate the accuracy of the proposed approach, it is compared with four well-known neural classifiers from the literature.