ICRA 2021：Self-Supervised Person Detection in 2D Range Data using a Calibrated Camera（SSPD）

下方↓公众号后台回复“SSPD”，即可获得论文电子资源。

现有的问题：
only a few annotated datasets are available for training and testing these deep networks, potentially limiting their perfor- mance when deployed in new environments or with different LiDAR models. （只有少数带注释的数据集可用来训练和测试这些深度网络，这可能会限制它们在新环境或不同LiDAR模型下的性能。）

• While early approaches for detecting persons in 2D range data focused on heuristics with hand-crafted features [1], [2], recent studies used convolutional neural networks and further improved the detection results [3], [4]（早期在2D范围数据中检测人的方法主要是使用手工制作的特征[1]，[2]的启发式方法，而最近的研究使用了卷积神经网络，并进一步改进了检测结果[3]，[4]）

文章工作：

• We propose a method, which uses bounding boxes from an image-based detector (e.g. Faster R-CNN) on a calibrated camera to automatically generate training labels (called pseudo-labels) for 2D LiDAR-based person detectors.（我们提出了一种方法，它使用标定摄像机上基于图像的检测器(例如Faster R-CNN)的边界框来自动生成基于2D lidar的人检测器的训练标签(称为伪标签)。）
• we proposed a method to automatically generate pseudo-labels for training 2D LiDAR-based person detectors, using bounding boxes generated from an image- based person detector with a calibrated camera.（我们提出了一种自动生成伪标签的方法，用于训练基于二维激光雷达的人探测器，使用边界框从基于图像的人探测器与标定摄像机生成），Even stronger detec- tors were obtained by combining pseudo-labels with robust training techniques.（通过将伪标记与鲁棒训练技术相结合，得到了更强的检测器。）
• we propose a method to automatically generate labels for training LiDAR-based person detectors using the output of an image-based detector on a calibrated camera（我们提出了一种方法，使用一个基于图像的探测器输出在一个校准的相机上自动生成标签训练基于激光雷达的人探测器）

对工作的评价

• Our method is an effective way to improve person detectors during deployment without any additional labeling effort, and we release our source code to support relevant robotic applications.(我们的方法是一种在部署过程中改进人员检测器的有效方法，无需任何额外的标记工作，我们发布了源代码来支持相关的机器人应用程序。）

• Our method provides an effective way to bridge the do-main gap between data encountered during training and dur- ing deployment. （我们的方法提供了一种有效的方法来弥合在训练和部署过程中遇到的数据之间的鸿沟。）

Introduction

whereas relatively little attention has been given to annotating 2D range data (see Fig. 1).（而对二维距离数据的注释则相对较少(见图1)。）

We utilize the image-based annotations to train 2D LiDAR-based person detectors by generating pseudo-labels from image-based detections.

main contributions of this work are：（主要贡献）

• We propose a method to automatically generate pseudo- labels for training 2D LiDAR-based person detectors, leveraging the output of an image-based detector and known extrinsic camera calibration.（利用基于图像的检测器的输出和已知的外部摄像机校准。）
• We demonstrate that the generated pseudo-labels can be used to train or fine-tune a person detector and experiment with robust training techniques to further improve its performance.
• We release our code, implemented in PyTorch with an easy-to-use detector ROS node, for robotic applications.

II. RELATED WORK

• A. LiDAR-based Person Detection
• B. Automatic Label Generation
• C. Learning with Noisy Labels

III. GENERATING PSEUDO-LABELS

We first use an object detector (e.g. Faster R-CNN [9]) to obtain person bounding boxes.
From all bounding boxes, a subset is selected using the following constraints:

• classification score greater than a threshold Tc,
• aspect ratio, the ratio between width and height, smaller than TAR,
• overlap ratio, with any other bounding box is smaller than To. The overlap ratio is defined as the intersection area divided by the area of the box.

This proposed method assumes that the person belongs to the foreground of the scene and is the dominant object in the cropped LiDAR scan, which is typically satisfied by the content of detection bounding boxes.（该方法假设人属于场景的前景，是裁剪激光雷达扫描的主要目标，通常由检测边界框的内容来满足。）

IV. PERSON DETECTION WITH PSEUDO-LABELS