Video prediction method, video prediction apparatus, electronic device and vehicle

Abstract

Disclosed are a video prediction method, a video prediction device, electronic equipment and a vehicle. The video prediction method includes: a training step, using the posterior probability distribution generated by the posterior encoder as the ground truth, to train the prior encoder; and a prediction step, using the prior encoder to generate the prior probability distribution from known frames , using the prior probability distribution as a latent variable for the first decoder, which generates predicted frames from known frames. Therefore, the priori encoder can be trained using the posterior probability distribution generated by the posteriori encoder, and then the priori distribution of the priori encoder can be used as a hidden variable for video prediction, so that for random video prediction problems, a more robust A priori estimate of the hidden variable of .

Description

technical field [0001] The present disclosure generally relates to the technical field of automatic driving, and specifically relates to a video prediction method, a video prediction device, electronic equipment and a vehicle. Background technique [0002] In recent years, autonomous driving or advanced driver assistance systems (ADAS) have received increasing attention. ADAS systems usually need to use various on-board sensors to perceive various states of the vehicle itself and the surrounding environment, and then identify, detect and track static and / or dynamic objects based on the collected data, and combine map data to perform calculations and analysis, so as to formulate driving strategies, and finally realize the automatic driving function. [0003] In automatic driving scenarios, it is usually necessary to predict the movement of dynamic objects (such as pedestrians, vehicles, etc.) in the environment based on video or image sequences obtained through image acquisi...

AI Technical Summary

Problems solved by technology

However, these methods are difficult to cope with real-time requirements due to low computational efficiency, or poor prediction results, or are not robust

Images