CN-115169677-B - Pedestrian long-term track prediction method and device

CN115169677BCN 115169677 BCN115169677 BCN 115169677BCN-115169677-B

Abstract

The invention provides a pedestrian track prediction method, a device and a medium, wherein the method comprises the steps of obtaining a pedestrian historical position map, a global scene map and a pedestrian historical observation track; the method comprises the steps of obtaining a pedestrian scene position feature according to the pedestrian history position map and the global scene map, modeling the pedestrian history observation track to generate a pedestrian future movement trend, predicting a pedestrian terminal according to the pedestrian future movement trend and the pedestrian scene position feature, and predicting a future track by combining the pedestrian terminal. The method combines the future motion trend of the pedestrian, the global scene layout and the long-term track prediction model of the multi-mode target, and improves the prediction accuracy on the long-term track prediction of the pedestrian.

Inventors

MAO TIANLU
WANG YISU
WANG ZHAOQI

Assignees

中国科学院计算技术研究所

Dates

Publication Date: 20260505
Application Date: 20220630

Claims (7)

1. A pedestrian long-term trajectory prediction method, characterized by comprising: acquiring a pedestrian history position map, a global scene map and a pedestrian history observation track; Obtaining scene position features of pedestrians according to the historical position map and the global scene map of the pedestrians, wherein the scene position features of the pedestrians comprise first scene position features related to destination targets and second scene position features related to tracks; Modeling the historical observation track of the pedestrian to generate a future motion trend of the pedestrian; predicting the pedestrian track according to the future movement trend of the pedestrian and the scene position characteristic of the pedestrian comprises the following steps: obtaining a pedestrian end point target map according to the first scene position characteristics; Dividing each region in the global scene map into three classification forms of a frequent traffic region, an occasional traffic region and an forbidden traffic region, and constructing a scene layout correction map according to the probability of passing pedestrians in the three classification traffic regions in a probability form from large to small; Correcting the pedestrian end point target map through the scene layout correction map, discarding the pedestrian end point target of the no-traffic area, and obtaining a corrected pedestrian end point target map, wherein the corrected pedestrian end point target map comprises a pedestrian end point target of a frequent traffic area and a pedestrian end point target of an occasional traffic area; Downsampling the future movement trend of the pedestrian and the corrected pedestrian end point target map to enable the future movement trend of the pedestrian and the pedestrian end point target map to correspond to the dimensions of each layer of the scene position code one by one; And merging the corrected pedestrian end point target map, the pedestrian future motion trend and the second scene position characteristic into each layer of the track decoder, and decoding to obtain the pedestrian track.
2. The method of claim 1, wherein the obtaining scene location features of the pedestrian from the pedestrian historic location map and the global scene map comprises: splicing the pedestrian history position map and the global scene map to be used as the input of a pedestrian scene position encoder, And encoding the scene position by using a pedestrian scene position encoder to obtain scene position characteristics of pedestrians.
3. The method of claim 1, wherein modeling the historical observation trajectory of the pedestrian to generate a future motion trend of the pedestrian comprises: the multilayer perceptron is utilized to carry out trend encoding and decoding on the historical observation track of the pedestrian, so as to generate future motion trend of the pedestrian; the future movement trend of the pedestrian is expressed as: Wherein, the A motion trend encoder representing the composition of the multi-layer perceptron, For a historical observation of the trajectory of a pedestrian, Parameters representing the motion trend encoder, A motion trend decoder representing the multi-layer perceptron, Parameters representing the motion trend decoder, Representing the resulting future movement trend of the pedestrian.
4. The method of claim 1, wherein the step of determining the position of the substrate comprises, Extracting the first scene location feature and the second scene location feature from the scene location features of the pedestrian using an attention mechanism.
5. The method of claim 1, wherein the step of determining the position of the substrate comprises, The pedestrian end point target map is expressed as: Wherein, the Representing the generated pedestrian end-point target map, Representing the end-point target decoder based on the convolutional layer, Parameters representing the destination target decoder of the endpoint, Representing a first scene location feature associated with the endpoint target.
6. The method of claim 1, wherein the step of determining the position of the substrate comprises, The pedestrian trajectory is expressed as: Wherein, the Representing a second scene location feature As a trend of future movement of the pedestrian, For a pedestrian end point target map, The operation of the splice is indicated and, In the case of a track decoder, As a parameter of the track decoder, Is a pedestrian trajectory.
7. A pedestrian long-term trajectory prediction apparatus, characterized in that the pedestrian long-term trajectory prediction method according to any one of claims 1 to 6 is employed, the apparatus comprising at least: the acquisition module is used for acquiring a pedestrian history position map, a global scene map and a pedestrian history observation track; the scene position coding module is used for obtaining scene position features of pedestrians according to the historical position map and the global scene map of the pedestrians, wherein the scene position features of the pedestrians comprise first scene position features related to the destination targets and second scene position features related to the tracks; The pedestrian movement trend modeling module is used for modeling the historical observation track of the pedestrian to generate a future movement trend of the pedestrian; The pedestrian track generation module predicts a pedestrian track according to the future motion trend of the pedestrian and the scene position characteristic of the pedestrian, and comprises the following steps: obtaining a pedestrian end point target map according to the first scene position characteristics; Dividing each region in the global scene map into three classification forms of a frequent traffic region, an occasional traffic region and an forbidden traffic region, and constructing a scene layout correction map according to the probability of passing pedestrians in the three classification traffic regions in a probability form from large to small; Correcting the pedestrian end point target map through the scene layout correction map, discarding the pedestrian end point target of the no-traffic area, and obtaining a corrected pedestrian end point target map, wherein the corrected pedestrian end point target map comprises a pedestrian end point target of a frequent traffic area and a pedestrian end point target of an occasional traffic area; Downsampling the future movement trend of the pedestrian and the corrected pedestrian end point target map to enable the future movement trend of the pedestrian and the pedestrian end point target map to correspond to the dimensions of each layer of the scene position code one by one; And merging the corrected pedestrian end point target map, the pedestrian future motion trend and the second scene position characteristic into each layer of the track decoder, and decoding to obtain the pedestrian track.

Description

Pedestrian long-term track prediction method and device Technical Field The invention relates to the technical field of crowd track prediction, in particular to a method and a device for predicting long-term tracks of pedestrians. Background Most of the existing pedestrian track prediction methods are used for researching the problem of short-time track prediction, the prediction time related to long-time track prediction is longer, the moving distance of pedestrians is longer, the surrounding scene changes are larger, and the influence of short-term pedestrian interaction in the observation stage on the future long-term motion of the pedestrians is gradually weakened. The target position of the pedestrian is used as the embodiment of the pedestrian movement intention, so that the general trend of the pedestrian movement is determined, and the influence on the future route of the pedestrian is gradually increased. Most of the existing pedestrian track prediction methods are used for researching the short-time track prediction problem, and the prior short-time track prediction has the defects of short pedestrian movement distance, few related scene changes and certain limitation in practical application. The long-term track prediction relates to longer prediction time, the movement distance of the pedestrians is longer, the surrounding scene changes are larger, the influence of short-term pedestrian interaction on the future long-term movement of the pedestrians in the observation stage is gradually weakened, the target positions of the pedestrians are used as the representation of the movement intention of the pedestrians, and the influence on the future routes of the pedestrians is gradually increased. Although the conventional Ynet model realizes long-term track prediction, the conventional Ynet model still has defects in association modeling of people and scenes and extraction of pedestrian end points and pedestrian motion modes by using scene layout, so that predicted targets can be positioned at positions such as trees, buildings and the like or the situation that future tracks of pedestrians are inconsistent with the motion trend of the pedestrians exists. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a long-term track prediction method and device for pedestrians, and the method combines the future motion trend of the pedestrians, global scene layout and a long-term track prediction model of a multi-mode target, so that the prediction accuracy is improved on the long-term track prediction of the pedestrians. In order to achieve the above object, an aspect of the present invention provides a pedestrian long-term trajectory prediction method, including: acquiring a pedestrian history position map, a global scene map and a pedestrian history observation track; obtaining scene position features of pedestrians according to the historical position map and the global scene map of the pedestrians; Modeling the historical observation track of the pedestrian to generate a future motion trend of the pedestrian; Predicting the pedestrian track according to the future movement trend of the pedestrian and the scene position characteristic of the pedestrian. Optionally, the obtaining the scene position feature of the pedestrian according to the historical position map and the global scene map of the pedestrian includes: splicing the pedestrian history position map and the global scene map to be used as the input of a pedestrian scene position encoder, And encoding the scene position by using a pedestrian scene position encoder to obtain scene position characteristics of pedestrians. Optionally, the modeling the historical observation track of the pedestrian to generate a future motion trend of the pedestrian includes: the multilayer perceptron is utilized to carry out trend encoding and decoding on the historical observation track of the pedestrian, so as to generate future motion trend of the pedestrian; the future movement trend of the pedestrian is expressed as: Xtend＝MLPdec(MLPenc(Xobs;Wme);Wmd) Wherein, MLP enc represents a motion trend encoder formed by a multi-layer perceptron, X obs represents a historical observation track of a pedestrian, W me represents parameters of the motion trend encoder, MLP dec represents a motion trend decoder formed by the multi-layer perceptron, W md represents parameters of the motion trend decoder, and X tend represents a finally generated future motion trend of the pedestrian. Optionally, the scene location features of the pedestrian include a first scene location feature associated with the end point target and a second scene location feature associated with the trajectory; the predicting the pedestrian track according to the future movement trend of the pedestrian and the scene position characteristic of the pedestrian comprises the following steps: obtaining a pedestrian end point target map according to the first scene position char