Attention mechanism-based pedestrian trajectory prediction method

Abstract

The invention relates to an attention-based pedestrian trajectory prediction method, which is used for more accurately and quickly predicting a future trajectory of a pedestrian. The invention specifically comprises three modules: an individual attention coding module used for calculating similarity of hidden vectors in a historical trajectory of a pedestrian and outputting an individual attention feature matrix so as to obtain main influence factors of the pedestrian in a motion process; the social attention pooling module that is used for receiving a calculation result, namely an individual attention feature matrix, of the individual attention coding module, calculating the similarity of hidden vectors in historical tracks of all pedestrians in a scene and outputting a comprehensive motion feature matrix so as to obtain a mutual influence relationship among the pedestrians in a motion process; the gating circulation unit decoding module that is used for receiving a calculation result of the social attention pooling module, namely a comprehensive motion characteristic matrix, and calculating and outputting future trajectory coordinates of pedestrians by utilizing a gating circulation unit. The method effectively improves the prediction precision and speed.

Description

technical field [0001] The invention relates to the fields of computer vision technology and automatic driving, and is a pedestrian trajectory prediction method based on an attention mechanism. Background technique [0002] Pedestrian trajectory prediction is one of the important research directions in computer vision applications. The research results can be widely used in pedestrian avoidance, automatic navigation, street planning, automatic driving and other fields. The trajectory prediction problem can be regarded as a sequential problem, that is, to predict the future trajectory of pedestrians based on their historical trajectories in the scene. Because the pedestrian motion is flexible and the interactive motion between pedestrians is complex and abstract, the main challenge of pedestrian trajectory prediction is how to accurately find the motion law of pedestrians and model the interactive motion between pedestrians. [0003] In pedestrian trajectory prediction, most...

AI Technical Summary

Problems solved by technology

[0004] Existing technical defects: the current method ignores the influence of pedestrian intentions on motion, resulting in the inability to correctly simulate the motion state of pedestrians; the current encoding module is limited to using a recurrent neural network to encode the position information of pedestrians at the current moment

However, the cyclic neural network contains a large number of repeated calculations, which leads to the long running time of the model, and the cyclic neural network has problems such as gradient disappearance and gradient explosion.

Images