Exploiting polar grid structure and object shadows for fast object detection in point clouds

Alsfasser, Martin; Siegemund, Jan; Kurian, Jittu; Kummert, Anton

doi:10.1117/12.2557043

Cited by 7 publications

(5 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The LiDAR data is analyzed in terms of both its spatial and temporal dimensions in order to create a point cloud that is derived from the recorded distances of reflection. The points can be represented using several coordinate systems, such as Cartesian coordinates (x, y, z) [46], [47], cylindrical coordinates (r, θ, z) [48], or spherical coordinates (θ, ϕ, r) [47]. In cylindrical and spherical coordinates, θ denotes the azimuth angle in the x,y plane.…”

Section: Related Workmentioning

confidence: 99%

Utilizing a Single-Stage 2D Detector in 3D LiDAR Point Cloud With Vertical Cylindrical Coordinate Projection for Human Identification

Eka Budiyanta,

Mulyanto Yuniarno,

Usagawa

et al. 2024

IEEE Access

View full text Add to dashboard Cite

Exploiting sensitive human data in human visual monitoring system violates individual's privacy. Hence, this study utilized a Light Detection and Ranging (LiDAR)-generated three-dimensional (3D) point cloud instead of an RGB camera as it captures the human object without detailed imagery. Given their dispersed nature, processing 3D LiDAR point cloud is economically inefficient as it requires some preliminary actions. Alternatively, this study applied a single-stage detection process using only one data type, namely 3D LiDAR point cloud with vertical axis direct projection. This approach utilized the 3D LiDAR cylindrical coordinates features to project the data onto two-dimensional (2D) image space with various computational capabilities and back-projection algorithm to restore the identified object on the 2D plane onto 3D LiDAR point cloud coordinates. This model also implemented the YOLOv5 series due to their varied sizes. The evaluation of this approach utilized accuracy metric which was the average of mean average precision (mAP) value based on different intersections over union (IoU) thresholds. The proposed methodology effectively employed a vertical projection technique to identify human objects. Notably, this approach distinguishes itself from previous methods such as PIXOR, BirdNet, BirdNet+, BevDetNet, and Frustum-PointPillars, offering a novel perspective in the field. In addition, the best and worst performing models had the accuracy values of 44.35% and 79.83%, with inference speeds of 3.7 ms and up to 25 ms, respectively. Further, the inference speeds of all models were less than 33.33 ms. Thus, the monitored objects were identified before the LiDAR system enters the next azimuth rotation.

show abstract

Section: Related Workmentioning

confidence: 99%

Utilizing a Single-Stage 2D Detector in 3D LiDAR Point Cloud With Vertical Cylindrical Coordinate Projection for Human Identification

Eka Budiyanta,

Mulyanto Yuniarno,

Usagawa

et al. 2024

IEEE Access

View full text Add to dashboard Cite

show abstract

“…To ensure physically possible locations the injected objects are compared in bird's-eyeview with other objects to avoid collisions [23]. The work of [2] additionally removed all points behind the cuboid of the injected object in polar coordinates to remove any possible overlap with existing objects, thus sidestepping the collision issue mentioned by [23]. In a more recent publication [11] applied a sub-sampling of the original sampled injection object along the scanlines and between scanlines to extend the distance of injections in the target point cloud, while keeping a similar structure of the lidar scanlines.…”

Section: B Local Augmentationsmentioning

confidence: 99%

What Can be Seen is What You Get: Structure Aware Point Cloud Augmentation

Hasecke

Alsfasser

Kummert

2022

2022 IEEE Intelligent Vehicles Symposium (IV)

Self Cite

View full text Add to dashboard Cite

To train a well performing neural network for semantic segmentation, it is crucial to have a large dataset with available ground truth for the network to generalize on unseen data. In this paper we present novel point cloud augmentation methods to artificially diversify a dataset. Our sensor-centric methods keep the data structure consistent with the lidar sensor capabilities. Due to these new methods, we are able to enrich low-value data with high-value instances, as well as create entirely new scenes. We validate our methods on multiple neural networks with the public SemanticKITTI [3] dataset and demonstrate that all networks improve compared to their respective baseline. In addition, we show that our methods enable the use of very small datasets, saving annotation time, training time and the associated costs.

show abstract

“…Some recent methods that operate on BEV start to explore polar voxels for point clouds. For 3D object detection, Alsfasser et al [1] voxelizes points under the Cylindrical Coordinate System, MVF [37] adopts both cuboid-shaped voxels and spherical voxels, and CVCNet combines cylindrical and spherical coordinate system into one Hybrid-Cylindrical-Spherical (HCS) coordinate system to detect object from both bird's eye view and range view. On the other hand, the success of PolarNet [35] and Cylinder3D [36] shows the advantage of Cylindrical grids over Cartesian voxels in LiDAR semantic segmentation.…”

Section: Related Workmentioning

confidence: 99%

“…In fact, polar grid outperforms the cartesian grid on the lidar segmentation task [35,36]. However, the detection peformance on a polar grid still lags the cartesian grid [1,5,22]. This is because of the distortion the objects undergo when this representation is ultimately unfolded to a rectangular representation to enable the use of convolutional layers.…”

Section: Introductionmentioning

confidence: 99%

PolarStream: Streaming Lidar Object Detection and Segmentation with Polar Pillars

Chen¹,

Vora²,

Beijbom³

2021

Preprint

View full text Add to dashboard Cite

Recent works recognized lidars as an inherently streaming data source and showed that the end-to-end latency of lidar perception models can be reduced significantly by operating on wedge-shaped point cloud sectors rather then the full point cloud. However, due to use of cartesian coordinate systems these methods represent the sectors as rectangular regions, wasting memory and compute. In this work we propose using a polar coordinate system and make two key improvements on this design. First, we increase the spatial context by using multi-scale padding from neighboring sectors: preceding sector from the current scan and/or the following sector from the past scan. Second, we improve the core polar convolutional architecture by introducing feature undistortion and range stratified convolutions. Experimental results on the nuScenes dataset show significant improvements over other streaming based methods. We also achieve comparable results to existing non-streaming methods but with lower latencies.

show abstract

Exploiting polar grid structure and object shadows for fast object detection in point clouds

Cited by 7 publications

References 0 publications

Utilizing a Single-Stage 2D Detector in 3D LiDAR Point Cloud With Vertical Cylindrical Coordinate Projection for Human Identification

Utilizing a Single-Stage 2D Detector in 3D LiDAR Point Cloud With Vertical Cylindrical Coordinate Projection for Human Identification

What Can be Seen is What You Get: Structure Aware Point Cloud Augmentation

PolarStream: Streaming Lidar Object Detection and Segmentation with Polar Pillars

Contact Info

Product

Resources

About