【作者】 彭涛；

【导师】 陆威；

【作者基本信息】 哈尔滨工业大学 ， 电子科学与技术， 2025， 硕士

【摘要】 盖革雪崩光电二极管(Geiger-mode avalanche photodiode,GM-APD)探测器凭借其单光子级探测灵敏度、纳秒级时间分辨率和超远探测距离,逐渐成为新型光电探测系统的研究热点。但是其视场角小,在扫描成像中因平台或视场运动引起的点云运动偏差,影响了三维重构精度。因此本文提出一种基于位姿校正的改进迭代最近点(Iterative Closest Point,ICP)动态点云配准方法,抑制点云运动畸变,获得宽视场、大范围、高品质三维点云图像。首先,针对载荷运动扫描场景下相邻帧间点云存在显著位姿漂移,传统ICP类算法对初始位姿敏感度高、在较大位姿偏差下易收敛至局部最优解的问题,提出一种基于位姿校正的点云粗配准方法,融合角度编码器、IMU和GPS,构造坐标变换公式,实现了三维点云到绝对位置的空间数据变换。经仿真和实验验证,在2km搜索扫描成像中平均距离误差为2.68m,在400m摆动扫描成像中平均距离误差为0.48m。其次,针对载荷运动扫描场景下相邻帧间点云重叠率低,且易强制配准非匹配点对的问题,提出基于KD-Tree的双向匹配机制提取重叠点云,再基于多分辨率邻域优化改进ICP实现点云精配准。仿真和实验表明,在2km搜索扫描成像中平均距离误差可达到0.21m,在400m摆动扫描成像中平均距离误差可达到0.10m。最后,针对点云中存在大量离群噪声,且主轴方向与坐标轴存在显著夹角的问题,采用双阈值判定改进统计滤波滤除离群点,降低点云噪声水平,再基于主成分分析(Principal Component Analysis,PCA)和旋转卡壳法自主识别点云方向,通过旋转矩阵将点云主轴方向与坐标轴对齐。仿真和实验结果显示,双阈值判定统计滤波在密集噪声中可有效滤除离群点。在2km仿真成像点云中,经过方向识别后生成的距离像的有效像素占比高达81.36%。在400m真实成像点云中,有效像素占比可达58.38%。这将为后续跟踪、识别等任务提供高质量的输入,大大节省存储空间,并提高计算效率。更多还原

【Abstract】 The Geiger-mode avalanche photodiode(GM-APD)detector,with its single-photon-level detection sensitivity,nanosecond time resolution,and ultra-long detection range,has gradually become a research hotspot in novel optoelectronic detection systems.However,its small field of view and point cloud motion deviations caused by platform or field movement during scanning imaging affect the accuracy of 3D reconstruction.Therefore,this paper proposes an improved ICP dynamic point cloud registration method based on pose correction,which suppresses point cloud motion distortions and achieves wide-field,large-scale,and high-quality 3D point cloud images.Firstly,to address the significant pose drift between adjacent frames in payload motion scanning scenarios,where traditional ICP algorithms are highly sensitive to initial poses and prone to converging to local optima under large pose deviations,a coarse point cloud registration method based on pose correction is proposed.This method integrates angle encoders,IMU,and GPS to construct a coordinate transformation formula,enabling the spatial data transformation of 3D point clouds to absolute positions.Simulation and experimental verification show that the average distance error in 2km search scanning imaging is 2.68m,and in 400m swing scanning imaging,it is 0.48m.Secondly,to tackle the issue of low overlap rates between adjacent frames and the tendency to force registration of non-matching point pairs in payload motion scanning scenarios,a KD-Tree-based bidirectional matching mechanism is proposed to extract overlapping point clouds,followed by ICP improvement through multi-resolution neighborhood optimization for fine point cloud registration.Simulations and experiments demonstrate that the average distance error in 2km search scanning imaging can reach 0.21m,and in 400m swing scanning imaging,it can reach 0.10m.Finally,to address the presence of numerous outlier noises in the point cloud and the significant angle between the principal axis and the coordinate axis,a double-threshold statistical filtering method is employed to remove outliers and reduce point cloud noise.Additionally,principal component analysis(PCA)and the rotating calipers method are used to autonomously identify the point cloud direction,aligning the principal axis with the coordinate axis through a rotation matrix.Simulation and experimental results show that the double-threshold statistical filtering effectively removes outliers in dense noise environments.In 2km simulated imaging point clouds,the effective pixel ratio of the generated range images reaches 81.36%after direction identification,while in 400m real imaging point clouds,it reaches 58.38%.This provides high-quality input for subsequent tracking,recognition,and other tasks,significantly saving storage space and improving computational efficiency.更多还原