Pedestrian trajectory prediction method based on graph partition convolutional neural network (GP-CNN)

Abstract

The invention relates to a pedestrian trajectory prediction method based on a GP-CNN, and aims to solve the core problem of trajectory prediction in automatic driving, namely how to design a model to better capture associated interaction information to improve prediction precision and safety of an automatic driving automobile. According to the invention, a pedestrian trajectory prediction model applied to a complex scene is designed, based on the GP-CNN, interaction features of the scene are extracted by using a mode of combining two channels to serve as input, and information of time-domain features of the pedestrian trajectory is specially processed, meanwhile, forward and backward propagation of the prediction trajectory is smoother through residual connection, and then a pedestrian interaction prediction trajectory is generated through a trajectory prediction network.

Description

technical field [0001] The present invention relates to a technology in the field of unmanned driving forecasting and planning, in particular to a pedestrian trajectory prediction technology based on a graph-partitioned convolutional neural network. Background technique [0002] Autonomous driving will be closely related to people's lives in the future, so the goal of autonomous driving is to be implemented in real life, and no matter what aspect of real life, it is full of various scenarios. In other words, the driving scene in which the vehicle is located is a highly unknown and multi-burst, uncertain environment. Uncertainties may be sensor limitations due to occlusion and limited sensor range, from probabilistic predictions of other vehicles, from unknown social behavior in new areas. Under these uncertain conditions, in order to drive safely and efficiently, a predictive module for autonomous driving should intelligently utilize all available information and properly r...

AI Technical Summary

Problems solved by technology

Compared with the existing scene interaction feature extraction method, this method adopts the feature extraction method of combining the embedded graph partition channel and the graph convolution channel, and explicitly learns the behavior interaction weight in the scene through the graph partition adaptive, and at the same time Combined with TCN to extract time-domain interaction features, it solves the problem of direct learning of interaction relations in the time domain

Images