CS-Wild-Places Dataset

CS-Wild-Places is a large-scale lidar dataset for cross-source place recognition between ground and aerial viewpoints in forest environments. The data was collected from four forests in Brisbane, Australia over ten months, with lidar scans captured by a handheld sensor payload (below canopy) and aerial drone (above canopy). CS-Wild-Places builds upon the Wild-Places dataset by introducing geo-registered aerial submaps covering Karawatha and Venman forests, enabling research into place recognition between challenging viewpoints. We further release ground and aerial data captured in two new forest environments: QCAT and Samford Ecological Research Facility. The aerial data spans 370 hectares in total, and we captured two ground sequences with 3.5km of total traversal, producing a total of ~36k high resolution lidar submaps with accurate 6-DoF poses and timestamps. We release the data in three main configurations: raw (submaps randomly downsampled to 500k points max), and post-processed (submaps voxel-downsampled with 0.8m voxels, ground points removed, with and without normalisation).\nLineage: The ground data was collected with a handheld sensor payload consisting of a VLP-16 lidar sensor spinning at a 45 degree angle to maximise field of view, an IMU, GPS antenna, and four cameras. For each collected sequence we use Wildcat SLAM to register the lidar data into a globally consistent map with accurate 6-DoF pose estimation, from which we produce our submaps by collecting points within a 2-second sliding window every 2 seconds along the sensor trajectory.\n\nThe aerial data is collected with two setups. For Karawatha, Venman, and QCAT, we used a DJI M300 quadcopter equipped with a VLP-32C lidar sensor. For Samford, we used an Acecore NOA hexacopter equipped with a RIEGL VUX-120 pushbroom lidar. Aerial global maps are geo-registered using Wildcat SLAM with GPS RTK, and further aligned with the ground global maps using ICP. Aerial submaps are uniformly sampled from a 10m-spaced grid spanning the aerial map. All submaps are stored in the local coordinate frame, and 6-DoF poses are stored in UTM coordinates.

CS-Wild-Places是一款面向森林环境下地面与空中视角跨源场所识别的大规模激光雷达（lidar）数据集。该数据集于澳大利亚布里斯班的四座森林中耗时十个月采集完成，分别通过林冠下方的手持传感器载荷与林冠上方的空中无人机获取激光雷达扫描数据。

CS-Wild-Places基于Wild-Places数据集拓展而来，新增了覆盖Karawatha与Venman森林的地理配准空中子地图（geo-registered aerial submaps），为挑战性视角下的场所识别研究提供支撑。我们还额外发布了两处全新森林环境中的地面与空中数据：QCAT森林与萨姆福德生态研究设施（Samford Ecological Research Facility）。

本次公开的空中数据总覆盖面积达370公顷，地面数据则采集了两条总行程3.5千米的序列，最终生成约3.6万个高精度激光雷达子地图，配套精准的6自由度（6-DoF）位姿与时间戳信息。数据集以三种主要配置发布：原始格式（子地图经随机下采样，单点云数量上限为50万），以及后处理格式（子地图采用0.8米体素下采样、移除地面点，包含归一化与未归一化两种变体）。

数据溯源
地面数据采用的手持传感器载荷包含以45度角旋转以最大化视场的VLP-16激光雷达、惯性测量单元（IMU）、GPS天线与四个摄像头。每条采集序列均通过Wildcat SLAM将激光雷达数据配准为全局一致的地图，并实现精准的6-DoF位姿估计；随后沿传感器轨迹每2秒采集一次2秒滑动窗口内的点云，从而生成对应的子地图。

空中数据采用两套采集方案：针对Karawatha、Venman与QCAT森林，我们使用搭载VLP-32C激光雷达的DJI M300四旋翼无人机；针对Samford森林，我们使用搭载RIEGL VUX-120推扫式激光雷达的Acecore NOA六旋翼无人机。空中全局地图通过搭载GPS RTK的Wildcat SLAM完成地理配准，并通过迭代最近点（ICP）算法与地面全局地图完成进一步对齐。空中子地图从覆盖空中地图的10米间距网格中均匀采样得到。所有子地图均以局部坐标系存储，6-DoF位姿则采用UTM坐标系存储。