【作者】 陈明；

【导师】 杨春平；

【作者基本信息】 电子科技大学 ， 工程硕士（专业学位）， 2020， 硕士

【摘要】 智能环境感知作为自动驾驶领域的重要组成部分,其目的在于通过收集传感器数据来对驾驶环境进行有效建模,而运动目标检测则是这一过程中最核心和关键的部分。在拥挤的交通环境对多个运动目标进行检测一直是一项极具挑战性的任务,面临着实际环境中复杂的场景结构、光照条件突变、阴影干扰和动态背景等诸多难题,上述挑战与困难使得自动驾驶中运动目标检测成为一项较难解决的问题。因此构建一个高鲁棒性、高精度的运动目标检测模型具有重要意义。本文结合激光雷达点云数据和视觉图像数据,基于深度学习网络和运动检测理论,开展了驾驶环境中的人车目标分类识别和运动检测算法研究,具体研究内容如下:(1)研究了激光雷达和图像传感器的原理,对采集的数据进行分析,针对点云数据稀疏无序、难以直接利用的特点,引入三角线性插值和随机采样一致性,提出了一种基于深度图的激光雷达点云稠密化及地面去除算法,将点云图像化并构建图像化点云数据集,保证了点云数据的有效性,并对数据的融合方式进行了对比和探讨。(2)研究了基于深度学习的目标检测算法。以自动驾驶感兴趣的人车等运动目标作为检测对象,针对驾驶环境场景结构复杂的问题并考虑到嵌入式平台计算性能的限制,本文结合点云数据和图像数据,设计了一个点云图像多尺度特征提取融合和多尺度预测的目标检测网络,以深度可分离卷积为基本特征提取单元,提出了基于注意力机制的特征融合及校正模块,在保证检测精度的前提下有效缩小了模型大小。(3)提出了基于两种数据融合的人车目标分类识别和运动评估算法。在人车目标检测网络的基础上,提出基于改进运动补偿的光流法,对自动驾驶环境下的人车目标的运动状态进行评估。通过在KITTI数据集上实验证明,多源数据融合模型的测试精度及鲁棒性都高于单数据的人车运动目标检测算法。更多还原

【Abstract】 Intelligent environment perception,the purpose of which is to model the driving environment by collecting sensor data effectively,is an important part of the field of autonomous driving.Moving object detection is the key part of this process.Detecting multiple moving targets in a crowded traffic environment has always been a challenging task,facing many difficult problems such as complex scene structure,sudden lighting conditions changes,shadow interference,and dynamic background in the actual environment.The above challenges and difficulties make moving object detection a difficult problem in autonomous driving.Therefore,it is of great significance to construct a highly robust and accurate moving object detection model.This thesis combines Lidar point cloud data and visual image data,based on deep learning networks and motion detection theory,to carry out research on human-vehicle target classification and recognition and motion detection in the driving environment.The specific research content is as follows:Firstly,the related theories of Lidar and camera sensor is studied,and the collected data are analyzed.Aiming at the sparse and disordered point cloud data,which is difficult to use directly,triangular linear interpolation and RANSAC algorithms are introduced.Based on depth map,a point cloud densification and ground removal algorithm is proposed.Then,a dataset which contains picture form of the point cloud is constructed,ensuring the validity of the point cloud data.The data fusion methods are compared and discussed.Secondly,the object detection algorithm based on deep learning is studied.Taking moving objects such as people and cars,which are interested in autonomous driving,as detection targets,in view of the complex structure of the driving environment and the limitations of the computing performance of the embedded platform,we combine point cloud data and image data to design a lightweight object detection network.This network is based on multi-scale feature fusion of point cloud and visual data,multi-scale prediction target detection module and deep separable convolution feature extraction module.A feature fusion and correction module based on the attention mechanism is also applied,which effectively reduces the model size on the premise of ensuring detection accuracy.Finally,a human-vehicle target classification and motion estimation algorithm based on two kinds of data fusion algorithm is proposed.Aiming at the dynamic background of the driving environment,based on the human-vehicle target detection network,improved motion compensation and optical flow methods are applied to evaluate the motion state of human-vehicle targets in an autonomous driving environment.Based on the experiments in the KITTI dataset,it is showed that the test accuracy and robustness of the multi-source data fusion model are higher than those of single data moving object detection algorithms.更多还原