Event prediction via spatio-temporal sequence analysis
Résumé
Event prediction often refers to the process of inferring dynamic information such as human pose change, object motion trajectory and event development direction from static images or partial video, which has broad application prospects in security monitoring, automatic driving, human-computer interaction and other fields. This article proposes the means of deep learning to conduct research on sequence images’ event prediction in time and space dimensions. To solve the problem of missing semantic information and lack of external information in the current event prediction research, we propose to adopt human skeleton constraint as guiding information, then complete the task from two aspects: skeleton detection and prediction in two-dimensional sequence images, image generation under skeleton guidance. The sequence images event prediction and generation model can generate realistic images with good continuity and perfect details on the basis of reasonable prediction in a short time. To verify the feasibility of our algorithm, we carried out a series of experiments on multiple deep learning network structures. The experimental results of the traditional motion detection dataset and the automatic driving dataset show the theoretical and practical value of the proposed algorithm in the field of sequence image event prediction, which has great potential for development.