Continuous pedestrian trajectory prediction method based on memory enhancement network

Abstract

The invention relates to the technical field of intelligent traffic, in particular to a continuous pedestrian trajectory prediction method based on a memory enhanced network, comprising an auto-encoder for encoding input information, and a controller for reading and writing memory from a key value memory; realizing multi-modal trajectory prediction by using a multi-hop attention mechanism, and generating a prediction trajectory through multi-hop iterative query; a future trajectory is reconstructed through a decoder, a training controller selects some typical key value memories from a current task and stores the typical key value memories in an external memory, and a part of previous data is played back every certain period in the current task. Current and previous key value memories are stored through an external memory, and in order to slow down catastrophic forgetting of a neural network, forgetting of a model to a previous task is slowed down by using a sparse experience playback method; for the uncertainty of future trajectories of pedestrians, a multi-hop attention mechanism is introduced to generate a reasonable multi-modal trajectory, so that the multi-modal trajectory is decoded and output.

Description

technical field [0001] The invention relates to the technical field of intelligent transportation, in particular to a method for predicting continuous pedestrian trajectories based on a memory-enhanced network. Background technique [0002] The prediction of pedestrian trajectories is the key to the safe and orderly operation of mobile intelligent bodies such as unmanned vehicles and robots. By observing and learning the past trajectories of surrounding pedestrians and the interaction between pedestrians, the agent can make a rough estimate of the pedestrian's future trajectory, so as to plan the driving path in advance and avoid conflicts with pedestrians. Different from only measuring the current position of the pedestrian, accurately predicting the future trajectory of the pedestrian can give the agent more reaction time and operation space, and ensure that the agent can make optimal decisions based on sufficient measurement data. [0003] Early pedestrian trajectory pre...

AI Technical Summary

Problems solved by technology

However, the lack of pedestrian location or surrounding environment information in these latent variables or random noise data may lead to inaccurate pedestrian motion trajectories generated finally

At the same time, due to the lack of external memory of the cyclic neural network, it may be difficult to process long-sequenced pedestrian sequences and cannot interact with external information.

[0004] In addition, most of the current pedestrian trajectory prediction models based on deep neural networks adopt the offline training mode, that is, training and verification are performed on a given single task; Encounter various complex scenes and sudden situations; if there is a significant difference in data distribution between the new scene and the old scene, the model will largely forget the knowledge learned in the old scene after adapting to the new scene data. The field of continuous learning is known as the "catastrophic forgetting problem"

Images