Finding next best views for autonomous UAV mapping through GPU-accelerated particle simulation

Adler, Benjamin; Xiao, Jing; Zhang, Jianwei

doi:10.1109/iros.2013.6696481

Cited by 4 publications

(6 citation statements)

References 10 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The proposed method not only combines the ORB algorithm and the ICP algorithm, but also filters the matching results in the process of 3D motion estimation. Motion modes include translation along the vector [1,0,0] T at the speed of 6cm/s, rotation around the vector [4,1,2] T at the speed of 60 ∘ /s, rotation around the vector [2,5,1] T at the speed of 20 ∘ /s and translation along the [1,0,0] T at the speed of 4cm/s. As can be seen from Table 3, the time consumption of method 1 is 7 to 9 times more than that of method 2 and the proposed method.…”

Section: Experiments Of 3d Motion Estimationmentioning

confidence: 99%

“…Fig. 8e is the depth image acquired in the next best view when visual object does rotation around the vector [2,1,1] T at the speed of 30 ∘ /s. Fig.…”

Section: Experiments Of Next Best Viewmentioning

confidence: 99%

“…In Table 4, N n is the number of surface points, N new is the number of new added points, R new is the new added rate. The motion mode 1 is that the visual object does translation along the vector [1,-1,-1] T at the speed of 2 ffiffi ffi 3 p cm=s, the motion mode 2 is that the visual object does rotation around the vector [2,1,1] T at the speed of 30 ∘ /s, the motion mode 3 is that the visual object does rotation around the vector [−4,1,2] T at the speed of 20 ∘ /s and translation along the vector [2,0,1] T at the speed of 2 ffiffi ffi 5 p cm=s. It can be seen that when the visual object is in different motions, the number of surface points N n , the number of new added points N new and the new added rates R new in the depth images acquired in different next best views which are calculated by the proposed method are almost same as these in the depth image acquired in the result of self-occlusion avoidance.…”

Section: Experiments Of Next Best Viewmentioning

confidence: 99%

“…Table 5 shows the time consumption of proposed method when visual objects are in different motions. The motion mode 1 is that the visual object does translation along the vector [1,-1,-1] T at the speed of 2 ffiffi ffi 3 p cm=s, the motion mode 2 is that the visual object does rotation around the vector [2,1,1] T at the speed of 30 ∘ /s, the motion mode 3 is that the visual object does rotation around the vector [−4,1,2] T at the speed of 20 ∘ /s and translation along the vector [2,0,1] T at the speed of 2 ffiffi ffi 5 p cm=s, the motion mode 4 is that the visual object does rotation around the vector [3,3,6] T at the speed of 40 ∘ /s and translation along the vector [0, 2,1] T at the speed of 2 ffiffi ffi 5 p cm=s, and the motion mode 5 is that the visual object does rotation around the vector [1,-2,2] T at the speed of 20 ∘ /s and translation along the vector [1,2,0] T at the speed of 2 ffiffi ffi 5 p cm=s. Table 5 shows that the average time consumption of proposed method is 100.51 ms.…”

Section: Experiments Of Next Best Viewmentioning

confidence: 99%

“…Li et al [20] extracted different views' features by unsupervised feature learning, and then trained classifiers to evaluate each view's discrimination ability to determine the next best view. Adler et al [1] sorted the candidate views by achievable information gain to determine the next best view. Mauro et al [24] proposed a next best view method based on the concept of view importance.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

A next best view method based on self-occlusion information in depth images for moving object

Zhang

et al. 2018

Multimed Tools Appl

View full text Add to dashboard Cite

The determination of next best view of a camera for moving object has wide application in dynamic object scenario, such as unmanned aerial vehicle and automatic recognition. The major challenge of this problem is how to determine the next best view while the visual object is moving. In this work, a novel next best view method based on selfocclusion information in depth images of moving object is proposed. Firstly, a depth image of moving object is acquired and self-occlusion detection is utilized in the acquired image. On this basis, the self-occlusion regions are modeled by utilizing space quadrilateral subdivision. Secondly, according to the modeling result, a method based on the idea of mean shift is proposed to calculate the result of self-occlusion avoidance corresponding to the current object. Thirdly, the second depth image of moving object is acquired, and the feature points in two images are detected and matched, then the 3D motion estimation is accomplished by the 3D coordinates of feature points which are matched. Finally, the next best view is determined by combining the result of self-occlusion avoidance and 3D motion estimation. Experimental results validate that the proposed method is feasible and has relatively high real-time performance.Multimed Tools Appl (2018) 77:9753-9777 https://doi

show abstract

Section: Experiments Of 3d Motion Estimationmentioning

confidence: 99%

“…Fig. 8e is the depth image acquired in the next best view when visual object does rotation around the vector [2,1,1] T at the speed of 30 ∘ /s. Fig.…”

Section: Experiments Of Next Best Viewmentioning

confidence: 99%

Section: Experiments Of Next Best Viewmentioning

confidence: 99%

Section: Experiments Of Next Best Viewmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

A next best view method based on self-occlusion information in depth images for moving object

Zhang

et al. 2018

Multimed Tools Appl

View full text Add to dashboard Cite

show abstract

Autonomous Exploration of Urban Environments using Unmanned Aerial Vehicles

Adler¹,

Xiao

Zhang³

2014

Journal of Field Robotics

Self Cite

View full text Add to dashboard Cite

This paper introduces a custom-built unmanned aerial vehicle, capable of autonomous exploration in urban environments. It consists of a multicopter, an inertial navigation system and two 2D laser range finders. In addition to a description of the hardware architecture and individual components being used, the authors also discuss challenges and problems that arose during its construction as well as optimizations and workarounds applied in the course of its development.Also presented is the software architecture, with a focus on a novel algorithm capable of generating multiple next best views (NBVs), sorted by achievable information gain. Although being designed for application on airborne platforms in urban environments, it works directly on raw point clouds and thus can be used with any sensor generating spatial occupancy information (e.g. LIDAR, RGBD-or time-of-flight-cameras). To satisfy constraints introduced by real-time operation on UAVs, the algorithm is implemented on a highly parallel SIMD architecture and benchmarked using GPUs from multiple hardware generations, using data from real flights. It is also compared against the previous, CPU-based proof of concept.As the underlying hardware imposes limitations with regards to memory access and concurrency, necessary data structures and further performance considerations are explained in detail.Open-source code for this paper is available at http://www.github.com/benadler/octocopter/.

show abstract