A method for highway video semantic segmentation and related equipment

Abstract

The application provides a highway video semantic segmentation method and related equipment, and relates to the technical field of computer vision. Through super-resolution reconstruction on the obtained low-resolution highway video, the application obtains a to-be-segmented highway video in a target resolution state, and calls a multi-modal large language model to perform attention fusion denoising on the key video frame sequence of the to-be-segmented highway video and the obtained target segmentation object text description information, to obtain a target positioning attention map sequence for the target segmentation object. Then, segmentation mask track generation is performed based on the target positioning attention map sequence, and track propagation is performed on the generated target segmentation mask track set in the to-be-segmented highway video, to obtain a target semantic segmentation result of the to-be-segmented highway video about the target segmentation object, so that the fine semantic segmentation effect of a complex highway video is quickly realized.

Images