可导局部栅格化的端到端点云目标检测方法研究

点云目标检测是面向自主无人系统的智能空间感知和场景理解的关键技术之一。传统的点云目标检测存在点云表征学习困难、无法克服显著类内差异、检测流程繁琐等科学难题，难以在复杂场景中得到大规模应用。本项目针对以上难题，拟开展可导局部栅格化的端到端点云目标检测方法研究，实现高精度、高鲁棒性、低复杂度的三维目标识别与定位。主要研究内容：构建一种新型可导局部栅格化的点云表征学习架构，研究区域化小型卷积神经网络方法，实现高效的点云精细化特征提取；设计一种新型尺度感知目标识别与定位策略，建立自动感知目标尺度和自适应调整子网络权重的动态检测头网络，提升检测器的鲁棒性；研究分类定位联合损失下的一对一训练样本分配方法，构建对待检目标直接输出局部最优检测结果的模型，实现无需非极大值抑制的端到端检测。本项目研究成果为复杂场景下的点云目标检测奠定理论和技术基础，进而推动自主无人系统的智能空间感知和场景理解的发展和应用。

3D object detection in point clouds servers as a fundamental step towards intelligent spatial perception and scene understanding in autonomous unmanned systems. However, existing detection methods in point clouds fail to be applied massively in practical applications due to limited representational learning capability, low robustness and complex detection processes. This program aims to develop an accurate, robust and fast framework for end-to-end detection based on local grid rendering networks. The study includes: propose a novel backbone for point cloud feature extraction with simple design and high efficiency, which renders the small neighborhood of a subset of input points into a low-resolution 3D grid independently, allowing small-size CNNs to accurately exploit informative spatial structures contained in local regions efficiently; develop a scale-aware object recognition and localization strategy, which automatically determines the scale of objects and adaptively adjusts the weights of multiple built-in sub-networks, forming a dynamic detection head that is shown to be robust to large variance in instance scales; investigate a one-to-one label assignment strategy by considering both classification and localization cost between sample and ground-truth, such that the network predicts one single optimal result for each target during testing, achieving end-to-end detection without non-maximum suppression process. This program aims to provide new and critical technical support for intelligent spatial perception and scene understanding in complex scenes, thus promoting the development and application of intelligence technology for autonomous unmanned systems.