Gait Pattern Recognition via Fusing Static and Dynamic Features in Single Stream Network

Isha Jain, Navya, Dharini Reddy, Vijay Kumar Pandey, Tushar Sandhan

Abstract: Gait refers to the walking and motion characteristics of an individual. In this paper, we suggest a unique method for analyzing human gait patterns using static as well as dynamic features passing through the same convolutional neural network. Generally, the gaits are processed using contour detection and segmentation to perform time series analysis with extracted features. Furthermore, machine learning models such as decision trees, support vector machine are also utilized for the gait classification. We initially trained our model on both cropped average images and optical flow of each gait independently. Subsequently, we fused the average image of each gait (building static information) with its corresponding optical flow (dynamic information). This novel fusion approach facilitated the training of a single stream convolutional neural network model. Our findings reveal a notable accuracy of ${9 8. 7 5 \%}$ in successfully recognizing individuals based on their gait patterns. Simplicity of the model and succinct fusion of static and dynamic features produce improved results with very less computational time.

External IDs:dblp:conf/icccnt/JainNRPS24