【作者】 张搏；

【导师】 余国义；

【作者基本信息】 华中科技大学 ， 软件工程， 2020， 硕士

【摘要】 激光雷达目标检测和跟踪是三维感知的核心环节,目标检测的准确度和目标跟踪的精确度共同决定了三维感知效果的上限。一方面,基于激光雷达鸟瞰图的目标检测方法开创性地将二维图像检测方法应用到三维目标检测,在激光雷达目标检测领域得到了广泛应用。然而,该方法在将点云投影成鸟瞰图的过程中,损失了点云高度信息和激光点之间三维空间位置信息,目标投影图像很容易和周围环境混淆,造成目标检测准确度低的问题。另一方面,基于扩展卡尔曼滤波的目标跟踪算法解决了线性卡尔曼滤波(KF)不能应用于非线性系统的问题,但是该算法需要求解雅克比矩阵,计算量很大,在采用线性的方式对非线性系统进行近似逼近时,损失了高阶项,造成跟踪精确度低的问题。以上两个方面的问题已经成为制约激光雷达感知效果提升的主要瓶颈。本文通过研究,综合分析了应用于激光雷达目标检测和跟踪的感知算法,在此基础上,设计了基于点云几何特征的卷积神经网络目标检测算法(G-Net),和基于无迹卡曼滤波(UKF)的目标跟踪算法(UKF-T)。其中,G-Net主要优点在于(1)通过特征提取网络自动将特征信息从点云的几何信息中提取出来,从而避免了点云的高度信息和空间位置信息的丢失;(2)使用2D卷积网络U-Net作为主干网络对提取出的特征进行编解码学习,避免使用3D卷积带来计算量大耗时多的问题。UKF-T主要优点在于(1)其计算量随状态维度线性增加,计算量少;(2)将无迹变换应用到卡尔曼滤波中,即通过生成关键点来模拟原系统状态的概率分布,达到了高阶近似效果。本文通过进行以上这些方面的改进提高了激光雷达的感知效果。本文改进的激光雷达目标检测算法G-Net,在KITTI数据集中3D汽车类别上,相对于基于鸟瞰图的目标检测算法,准确度提升了24.18%;本文提出的目标跟踪算法UKF-T,在UDACITY数据集上相对于扩展卡尔曼滤波目标跟踪算法,精确度提升了39.63%;在真实的道路测试场景中,也能够准确检测到周围目标并进行持续稳定的跟踪,体现了本文算法的优越性。更多还原

【Abstract】 Lidar target detection and tracking is the core link of 3D sensing.The accuracy of target detection and the accuracy of target tracking jointly determine the upper limit of the 3D sensing effect.On the one hand,the target detection method based on the lidar bird’s eye view pioneeringly applies the two-dimensional image detection method to the three-dimensional target detection,which has been widely used in the field of lidar target detection.However,in the process of projecting a point cloud into a bird’s eye view,the method loses the height information of the point cloud and the three-dimensional spatial position information between the laser points.The projected image of the target is easily confused with the surrounding environment,resulting in low accuracy of target detection.On the other hand,the target tracking algorithm based on extended Kalman filter solves the problem that linear Kalman filter(KF)cannot be applied to nonlinear systems,but the algorithm requires solving the Jacobian matrix has a large amount of calculation.When the nonlinear method is used to approximate the nonlinear system in a linear manner,high-order terms are lost,causing a problem of low tracking accuracy.The above two problems have become the main bottlenecks that restrict the improvement of Li DAR sensing effect.Through research,this paper comprehensively analyzes perceptual algorithms applied to lidar target detection and tracking.Based on this,a convolutional neural network target detection algorithm(G-Net)based on point cloud geometric features is designed.Mann filter(UKF)target tracking algorithm(UKF-T).Among them,the main advantages of G-Net are(1)the feature information is automatically extracted from the geometric information of the point cloud through the feature extraction network,thereby avoiding the loss of the height information and spatial position information of the point cloud;(2)the use of 2D volumes Convolutional network U-Net is used as the backbone network to encode and decode the extracted features to avoid the problem of large computation and time consuming caused by3 D convolution.The main advantages of UKF-T are(1)the amount of calculation increases linearly with the state dimension,and the amount of calculation is small;(2)the unscented transformation is applied to Kalman filtering,that is,the probability distribution of the original system state is simulated by generating key points,A high-order approximation is achieved.This paper improves the perception effect of lidar by improving the above aspects.The improved lidar target detection algorithm G-Net in this paper improves the accuracy of the target detection algorithm based on the bird’s-eye view on the 3D car category in theKITTI dataset by an accuracy of 24.18%;the target tracking algorithm UKF-T proposed in this paper is in UDACITY Compared with the extended Kalman filter target tracking algorithm on the data set,the accuracy is improved by 39.63%.In real road test scenarios,the surrounding targets can also be accurately detected and continuously tracked,which shows the superiority of the algorithm in this paper.更多还原