Reducing Computational Cost in Pedestrian Conversation Activity Recognition through Skeleton Spatiotemporal Graphs

Le résumé fourni par la source

The existing method for recognizing conversation activity in walking pedestrian groups uses heavy three-dimensional mesh reconstruction and rendering, which significantly increases the computational cost. We investigate a lightweight framework that replaces skinned multi-person linear model mesh structures using skeleton spatiotemporal graphs extracted from general color video sequences to recognize the conversation activity labels. Compared with the existing method, our approach reduces inference time and graphics processing unit memory while maintaining equal or higher accuracy, which makes low computational cost conversation activity recognition feasible.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.