Emerging Gait Detection Based on Adaptive Weights of Human Body Parts

Gait detection is a crucial task in human identification, which can perform long-distance recognition without the need for subject cooperation, making it widely used in medical diagnostics and crime prevention. Existing identification methods suffer from reduced accuracy under varying identities and clothing conditions. Moreover, many gait detection models are inefficient, requiring excessive resources and time for training. This article introduces a deep network model, GaitAdapt, by adopting an adaptive human feature fusion method based on global and local strategies. Along with a dynamic data allocation approach, we can more effectively capture the spatiotemporal feature information of different body parts during pedestrian walking. Our method is resilient to disruptions in local features within the image and naturally enhances recognition accuracy by leveraging the motion trajectories of various body parts, even under complex clothing conditions. Experiments demonstrate that this model surpasses most mainstream baseline gait models of recent years in both short-term and long-term convergence speed and accuracy on popular public gait datasets, providing a new and effective approach for complex gait detection in current instrumentation systems.