Classification of Gait Patterns Using Overlapping Time Displacement of Batchwise Video Subclips

Abstract

Gait patterns of Cerebral Palsy (CP) patients have been used for cluster and classification analysis. Diplegia is the paralysis of one or more body parts which may be caused by CP and may come in various forms. Current clinical practice in gait issue diagnosis relies heavily on observation and is prone to human error. Following previous studies, the effectiveness of introducing modern machine learning techniques and processes in improving the classification accuracy on gait video data was investigated. This paper proposes a novel feature engineering approach by transforming the original video into overlapping sub-clips which not only maintains important features but also reduces training time. Multiple machine learning models have been constructed to examine their individual performances before ensembling them to improve overall performance. The ensemble architecture consists of two stages, a probabilistic-based aggregator and normalizer and a performance-weighted ensemble. Finally, the model classification accuracy was able to achieve over 95%, a marked improvement from the results obtained by the models applied on similar dataset from literature. Hence, this highlights the effectiveness of the proposed method in classification of gait patterns and potentially changing current clinical practice in gait-related diagnosis.