Abstract:

In order to solve the problems of the high cost of multi-channel video data collection and long-term high-quality annotation in high-rise buildings or complex open building environments, the generation of multi-channel video data across the field of view and the automatic annotation of pedestrian images was realized. Firstly, a virtual reality scene was designed to simulate pedestrian movement and automatically obtain marker data. Secondly, unsupervised domain adaptation methods were researched to reduce the difference in feature distribution between source and target domain data, enabling the model to generalize to the target building scene. Finally, the model's generalization ability was verified. Results show that the constructed virtual reality scene effectively overcomes the difficulties of cross-visual video data collection and high-quality annotation. The unsupervised domain adaptation method increased the average first hit rate from 22.02% to 45.48%. By combining source domain style conversion, data augmentation, and target domain pseudo label generation, the first hit rate has been increased by 20%, reducing distribution bias and achieving generalization of the model in different building scenarios.

Key words: building scenes, virtual simulation, cross-visual, personnel movement, automatic annotation