LocNet: Global Localization in 3D Point Clouds for Mobile Vehicles

Abstract

Global localization in 3D point clouds is a challenging problem of estimating the pose of vehicles without any prior knowledge. In this paper, a solution to this problem is presented by achieving place recognition and metric pose estimation in the global prior map. Specifically, we present a semi-handcrafted representation learning method for LiDAR point clouds using siamese LocNets, which states the place recognition problem to a similarity modeling problem. With the final learned representations by LocNet, a global localization framework with range-only observations is proposed. To demonstrate the performance and effectiveness of our global localization system, KITTI dataset is employed for comparison with other algorithms, and also on our long-time multi-session datasets for evaluation. The result shows that our system can achieve high accuracy.

OREOS: Oriented Recognition of 3D Point Clouds in Outdoor Scenarios

“…LocNet [6] use semi-handcrafted range histogram features as an input to a 2D CNN (Convolutional Neural Network), while Uy et al use a NetVLAD [20] layer on top of the PointNet [21] architecture [7]. Furthermore, Kim et al [22] recently presented the idea to transform point clouds into scan context images [23] and feed them into a CNN for sovling the place recognition problem.…”

“…In contrast to that, the two orientation specific vectors w A , and w S from the two close-by point clouds are fed into the Orientation Loss L θ . The combined loss L is then evaluated as described in Equation 6. We use ADAM [31] as a learning optimizer and use a learning rate of alpha = 0.001.…”

“…Efficient and descriptive data representations for place recognition using LiDAR point clouds remain, however, an open research question [3]- [5]. In contrast to our work, typical place recognition methods do not always explicitly deal with the full problem of estimating a 3 DoF transformation [6] [7] [8].…”

Semantically Assisted Loop Closure in SLAM Using NDT Histograms

“…Other methods make use of scan descriptors for the problem of global registration. LocNet [11] uses hand-crafted point cloud features that are then compared using a siamese neural network that compresses the features into the encoded representation. The features rely on the property of rotational lidar to generate concentric rings of points, and operate on each ring independently.…”

Have I Seen This Place Before? A Fast and Robust Loop Detection and Correction Method for 3D Lidar SLAM

“…As a result, many loops are not detected or too many false positives are present, which on its turn imposes a restriction on fast and reliable loop closure. More recently, researchers tend toward the application of convolutional neural networks (CNNs) to learn both the feature descriptors as well as the metric for matching them in a unified way [11,12,13,14]. A severe limitation of these methods on the other hand is that they need a tremendous amount of training data.…”