Publication

Abstract : Gait is the manner or pattern of walking governed by the coordination of muscles, joints, and the nervous system. Abnormal gait can affect general mobility and quality of life by stressing muscle joints and raising the chance of falling. Gait analysis poses several challenges, which include inter-class similarities and variations in walking patterns among individuals with the same pathology. The limited dataset and class imbalance further hinder the performance of the model. Earlier vision-based systems analyzing gait to detect these abnormalities were often limited to binary classification, distinguishing only between normal and abnormal. Later, some advanced systems attempted to identify specific pathologies. However, these models could not capture subtle variations in gait patterns, reducing overall performance. To address this issue, a novel network, cross reverse attention network (CRA-Net), which incorporates two complementary attention mechanisms called cross reverse attention block (CRAB) and cross attention block (CAB) to capture inter-class variations better, is proposed. CRAB adopts an inverse attention mask to extract complex features from discriminative regions and integrate them into features extracted by traditional Convolutional neural networks (CNN), making the model robust. CAB is designed to integrate diverse feature sets from local and global layers, thereby balancing multi-scale feature integration. Concatenating features from three levels of the proposed architecture makes the model learn all relevant features that aid in classifying gait pathologies. The experimental results have shown that CRA-Net effectively captures variations in gait patterns, achieving superior performance compared to existing methods.