【作者】 王海峰；

【导师】 桑海峰；

【作者基本信息】 沈阳工业大学 ， 工程硕士（专业学位）， 2022， 硕士

【摘要】 随着科学技术的发展,在监控场景下存在的智能系统数量越来越多,这些系统感知、理解进而预测人类移动的能力也变得越来越重要。行人轨迹预测技术在这种背景下应运而生,它在自动驾驶、服务机器人路径规划和智慧城市监控系统等领域都有着非常广泛的应用。传统的行人轨迹预测研究通常采用手工定义特征函数的方法,然而手工定义的特征很难模拟复杂的行人交互行为、计算复杂度较高且泛用性较差。近些年来,随着深度学习技术不断提升,基于复杂数据驱动的神经网络模型可以很好地解决上述的缺陷。因此,本文采用基于复杂数据驱动的神经网络模型的方法,在研究和分析现有行人轨迹预测模型的优缺点后,提出了两种行人轨迹预测模型。(1)针对现有模型中行人交互机制简单,无法确定每个行人不同的重要性,速度信息在行人交互时利用率不高的问题,提出一种基于注意力机制的行人轨迹预测模型。模型主体框架采用条件生成对抗网络,将显式速度信息作为条件标签调节网络生成,提高了速度信息对模型的影响。在生成器中的解码器部分设置注意力机制,在不同时刻为不同行人分配不同的注意力权重,增强了对行人交互信息的提取。判别器采用全卷积网络,分段对预测轨迹进行打分,使判别器分类效果提升。在ETH和UCY数据集上的实验结果表明,本文所提出的模型平均位移误差为0.48 m,最终位移误差为0.93 m。(2)针对在模型中缺乏关于行人交互时的时间和速度等信息建模的问题,提出一种基于图注意力机制的行人轨迹预测模型。模型整体架构同样采用速度标签的条件生成对抗网络,不同的是生成器中设计了基于增强图卷积注意力机制的时空信息融合模块,在提取行人轨迹序列运动特征并关注其空间上相互作用关系的同时,显式地编码行人序列的时间相关性。最后,将结合时空信息和速度信息的轨迹交互特征解码,完成轨迹的预测。此外,考虑到现有评价方法的不足,采用平均碰撞次数作为轨迹合理性的评判。在ETH和UCY数据集上进行实验验证,实验结果表明本文所提出的算法能有效地减少碰撞情况的发生,更好地完成行人轨迹预测,平均位移误差为0.40 m,最终位移误差为0.79 m。更多还原

【Abstract】 With the development of science and technology,there are more and more intelligent systems in the monitoring scene.The ability of these systems to perceive,understand and predict human movement has become more and more important.Pedestrian trajectory prediction technology came into being under this background.It is widely used in the fields of automatic driving,service robot path planning and smart city monitoring system.The traditional pedestrian trajectory prediction research usually adopts the method of manually defining the feature function.However,the manually defined features are difficult to simulate the complex pedestrian interaction behavior,high computational complexity and poor universality.In recent years,with the continuous improvement of deep learning technology,the neural network model based on complex data-driven can well solve the above defects.Therefore,this paper adopts the method of neural network model based on complex data-driven.After studying and analyzing the advantages and disadvantages of existing pedestrian trajectory prediction models,two pedestrian trajectory prediction models are proposed.(1)Aiming at the problem that the pedestrian interaction mechanism in the existing model is simple,the importance of each pedestrian can not be determined,and the speed information is not highly utilized in pedestrian interaction,a pedestrian trajectory prediction model based on attention mechanism is proposed.The main frame of the model adopts a conditional generative adversarial network,and uses the explicit speed information as a conditional label to regulate the network generation,which improves the influence of the speed information on the model.The attention mechanism is set in the decoder part of the generator,and different attention weights are assigned to different pedestrians at different times,which enhances the extraction of pedestrian interaction information.The discriminator uses a fully convolutional network to score the predicted trajectories in segments,so that the classification effect of the discriminator is improved.The experimental results on the ETH and UCY datasets show that the model proposed in this paper has an average displacement error of 0.48 m and a final displacement error of0.93 m.(2)Aiming at the lack of information about the time and speed of pedestrian interaction in the model,a pedestrian trajectory prediction model based on graph attention mechanism is proposed.The overall architecture of the model also adopts the conditional generative adversarial network of speed labels.The difference is that a spatial-temporal information fusion module based on the enhanced graph convolution attention mechanism is designed in the generator,which extracts the motion features of pedestrian trajectory sequences and pays attention to the spatial interaction between them.At the same time,the temporal correlation of pedestrian sequences is explicitly encoded.Finally,the trajectory interaction features combined with spatial-temporal information and speed information are decoded to complete trajectory prediction.In addition,considering the shortcomings of the existing evaluation methods,the average times of collisions is used as the evaluation of the reasonableness of the trajectory.Experiments are carried out on the ETH and UCY datasets.The experimental results show that the proposed algorithm can effectively reduce the occurrence of collisions and better complete the pedestrian trajectory prediction.The average displacement error is 0.40 m,and the final displacement error is 0.79 m.更多还原