Methods for Constructing and Augmenting Datasets of Violations in Power Operations Across Multiple Scenarios

Le résumé fourni par la source

The complexity of the power operation situation, the long-tailed distribution of the violation samples, and the difficulty in obtaining dynamic characteristics result in poor generalization ability of intelligent recognition models. In order to address this, in this paper, a method to construct and improve a dataset of power operation violation behaviors in multiple scenarios is proposed. First, there are three typical sites - substations, distribution rooms and towers erection - are selected. Multi-view videos are taken with the use of 360deg panoramic monitoring terminal. High-quality panoramic video frames are obtained using SURF feature registration, RANSAC estimation of homography matrix and Gaussian pyramid multi-resolution fusion. Second a hybrid data augmentation strategy is constructed. Simulated samples of semantic labels are created with the Unity 3D digital twins and a generator of StarGAN v2 is introduced based on a time shifting module to transfer the simulated samples to the real domains. The dataset built with this method enlarges the number of samples in rare categories such as "objects thrown from heights" from 98 to 3698; it shows the best performance in terms of average mAP (80.72%), which gives high-quality and balanced data support for the identification of power operation violations.

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