## Real-time Single-workstation Obstacle Avoidance Using Only Wide-field Flow Divergence (1996)

### Cached

### Download Links

Venue: | Journal of Computer Vision Research |

Citations: | 19 - 2 self |

### BibTeX

@INPROCEEDINGS{Camus96real-timesingle-workstation,

author = {Ted Camus and David Coombs and Martin Herman and Tsai-Hong Hong},

title = {Real-time Single-workstation Obstacle Avoidance Using Only Wide-field Flow Divergence},

booktitle = {Journal of Computer Vision Research},

year = {1996},

pages = {323--330}

}

### Years of Citing Articles

### OpenURL

### Abstract

This paper describes a real-time robot vision system which uses only the divergence of the optical flow field for both steering control and collision detection. The robot has wandered about the lab at 20 cm/s for as long as 26 minutes without collision.

### Citations

1260 |
Robust Regression and Outlier Detection
- Rousseeuw, Leroy
- 1987
(Show Context)
Citation Context ... to the rest of the data is unknown. The "breakpoint" of an estimator is defined as the smallest fraction (or percent) of contamination of a data set that can result in an arbitrarily incorr=-=ect value [32]-=-. In the case of least-sum-of-squares (a non-robust estimator), that percent tends to 0% for large (i.e., a single outlier can cause an arbitrarily incorrect result). In the case of the true median fi... |

560 |
Animate vision
- Ballard
- 1991
(Show Context)
Citation Context ...intelligent behavior by using minimalist visually-derived representations. In such representations, a minimal amount of information required to achieve the given task is extracted from the imagery [2]=-=[4]-=-. The representations contain only task-relevant information (i.e., relevant to obstacle avoidance), and the information is only represented in 2-D image coordinates. The control algorithms directly u... |

466 |
A computational framework and an algorithm for the measurement of visual motion
- Anandan
- 1989
(Show Context)
Citation Context ... problematic in our case where a wide-angle lens is used and individual objects occupy only a small fraction of the visual field. Calculating a least-squares best fit to the correlation surface as in =-=[3]-=- was ruled out due to real-time performance requirements (see efficiency experiments in [22][24]). In order to satisfy our real-time requirements, a fast approximation method was used to derive contin... |

410 |
Active vision
- Aloimonos, Weiss, et al.
- 1987
(Show Context)
Citation Context ...me intelligent behavior by using minimalist visually-derived representations. In such representations, a minimal amount of information required to achieve the given task is extracted from the imagery =-=[2]-=-[4]. The representations contain only task-relevant information (i.e., relevant to obstacle avoidance), and the information is only represented in 2-D image coordinates. The control algorithms directl... |

232 | Outline for a theory of intelligence
- ALBUS
- 1991
(Show Context)
Citation Context ... It is, therefore, simpler and faster. The obstacle avoidance system we describe in this paper is designed in accordance with the Real-Time Control System (RCS) hierarchical architecture described in =-=[1]-=-. RCS decomposes goals both spatially and temporally to meet system objectives. The system monitors its environment with sensors and updates models of the states of the system itself DRAFT, submitted ... |

151 |
Modern control systems
- Dorf, Bishop
- 1995
(Show Context)
Citation Context ...the robot's current heading , and from these the new target heading is calculated. (11) When a new desired heading is chosen, the robot steers smoothly to it with saturated negative feedback controls =-=[13]-=-. Desired change in heading, , is simply the difference between the desired heading and current heading. (12) y fq q' q' qfy++ = q' Dq Dq q' q -- = DRAFT, submitted 1997, revised 1999 23 The steering ... |

100 |
A fast two- dimensional median filtering algorithm
- Huang, Yang, et al.
- 1979
(Show Context)
Citation Context ...filters to the individual divergence values (process 3 in Figure 1). 5 Simple Robust Filters Median filtering performed on dense two-dimensional data can use fast running-histogram methods such as in =-=[19]-=- if the dynamic range of the data and desired quantization resolution of the median value can be specified. The algorithm described in [19] was intended for finding the true median and reduces an O(nm... |

95 | On the advantages of polar and log-polar mapping for direct estimation of timeto-impact from optical flow
- Tistarelli, Sandini
- 1991
(Show Context)
Citation Context ...-free (however crude) mobility as a base of competence, it is promising enough to be considered as a low-level competence in a goal-directed mobile robot system. 9 Discussion Some researchers [20][33]=-=[34]-=- have proposed using divergence or flow derivatives for visual cues, but did provide real-time implementations of these ideas. Nelson and Aloimonos [27][28] used directional flow divergence for stop-a... |

70 |
Surface orientation and time to contact from image divergence and deformation
- Cipolla, Blake
- 1992
(Show Context)
Citation Context ...y image coordinates is defined as: . (4) Note that (5) (6) where = . From equations (4) through (6), at : . (7) This can be rewritten in terms of the gradient as (8) (9) where is the surface gradient =-=[10]-=-. As stated, this equation holds for a planar coordinate system at . Because we are modeling our wide-angle camera using a spherical coordinate system, the spherical imaging surface at every point can... |

57 | Real-Time Obstacle Avoidance Using Central Flow Divergence, and Peripheral Flow
- Coombs, Herman, et al.
- 1998
(Show Context)
Citation Context ... the divergence data are already quantized to 8-bit values (necessary in ZT z Z nm nm DRAFT, submitted 1997, revised 1999 18 order to reduce the data bandwidth in the previous system's implementation =-=[11]-=-) so this extension is not necessary. Although the separable median filter is not guaranteed to find true median values, it reduces noise almost as well as a true median filter [26]. By definition, th... |

53 |
A parallel algorithm for real-time computation of optical flow
- Bulthoff, Little, et al.
- 1989
(Show Context)
Citation Context ...tation. (a) Test mobile robot (b) Robot's view of the lab through its 115-degree wide angle lens Figure 2 Robot and Its View DRAFT, submitted 1997, revised 1999 5 In correlation-based flow such as in =-=[5]-=-[17][21], the motion for the pixel at in one frame to a successive frame is defined to be the displacement of the patch of pixels centered at , out of possible displacements (where is an arbitrary par... |

53 |
Obstacle avoidance using flow field divergence
- Nelson, Aloimonos
- 1989
(Show Context)
Citation Context ...ntact, i.e., the time needed to reach an object at depth with velocity in the direction. A family of simple fixed flow divergence templates can be applied to any image sequence to estimate divergence =-=[28]-=-. Each template is symmetrically divided into positive and negative halves (Figure 9). Flow divergence is calculated by convolving the template with a window in the flow image and computing the sum of... |

36 | Accuracy vs. efficiency trade-offs in optical flow algorithms
- Liu, Hong, et al.
- 1998
(Show Context)
Citation Context ...er second on an 80 MHz Themis HyperSPARC 1 10 computer. Although this algorithm is not as accurate as many other optical flow algorithms, it is generally superior in terms of computational efficiency =-=[23]-=-. In our implementation, an original image of 256x512 pixels, captured with a field-of-view camera, is subsampled to 32x64 pixels using a simple block-subsampling algorithm which averages an NxN block... |

31 | Real-time quantized optical flow
- Camus
- 1997
(Show Context)
Citation Context ... system, it is important to have sufficient processor idle time available to buffer the overhead of the operating system. Otherwise, the image capture frame rate could vary from frame to frame. Camus =-=[7]-=-[8] describes a robust, real-time correlation-based optical flow algorithm which returns dense data even in areas of relatively low texture. This is the starting point of our implementation. (a) Test ... |

26 |
Inexact vision
- Thompson, Kearuey
- 1986
(Show Context)
Citation Context ...sion-free (however crude) mobility as a base of competence, it is promising enough to be considered as a low-level competence in a goal-directed mobile robot system. 9 Discussion Some researchers [20]=-=[33]-=-[34] have proposed using divergence or flow derivatives for visual cues, but did provide real-time implementations of these ideas. Nelson and Aloimonos [27][28] used directional flow divergence for st... |

24 | Robot navigation from a Gibsonian viewpoint
- Duchon, Warren
- 1994
(Show Context)
Citation Context ...mpling by a factor of 4 on a side is performed on the host workstation itself. The above algorithm returns quantized optical flow values. Although this is adequate for various robotic vision tasks [6]=-=[14]-=-[15][25], it is not sufficient for our application, because the calculation of divergence requires the ability to measure spatial derivatives of the optical flow. Because quantized optical flow is bas... |

22 | Real-Time Optical Flow
- Camus
- 1995
(Show Context)
Citation Context ...bsampling by a factor of 4 on a side is performed on the host workstation itself. The above algorithm returns quantized optical flow values. Although this is adequate for various robotic vision tasks =-=[6]-=-[14][15][25], it is not sufficient for our application, because the calculation of divergence requires the ability to measure spatial derivatives of the optical flow. Because quantized optical flow is... |

21 | A General Motion Model and Spatio-temporal Filters for 3-D Motion Interpretations
- Liu
- 1995
(Show Context)
Citation Context ... a small fraction of the visual field. Calculating a least-squares best fit to the correlation surface as in [3] was ruled out due to real-time performance requirements (see efficiency experiments in =-=[22]-=-[24]). In order to satisfy our real-time requirements, a fast approximation method was used to derive continuously-valued flow fields. To avoid a computationally expensive search for the true flow, th... |

16 | Ecological robotics: Controlling behavior with optical flow
- Duchon, Warren, et al.
- 1995
(Show Context)
Citation Context ...ng by a factor of 4 on a side is performed on the host workstation itself. The above algorithm returns quantized optical flow values. Although this is adequate for various robotic vision tasks [6][14]=-=[15]-=-[25], it is not sufficient for our application, because the calculation of divergence requires the ability to measure spatial derivatives of the optical flow. Because quantized optical flow is basical... |

16 |
A Separable Median Filter for Image Noise Smoothing
- Narendra
- 1981
(Show Context)
Citation Context ...intended for finding the true median and reduces an O(nm) complexity algorithm to approximatelysO(n) per pixel for an filtering window, where . It can, however, be generalized to the separable median =-=[26]-=- which approximates a two-dimensional true median filter by the successive application of two orthogonal one-dimensional median filters: first the median of each row of a two-dimensional window is com... |

12 |
A smart buffer for tracking using motion data
- Little, Kam
- 1993
(Show Context)
Citation Context ... (a) Test mobile robot (b) Robot's view of the lab through its 115-degree wide angle lens Figure 2 Robot and Its View DRAFT, submitted 1997, revised 1999 5 In correlation-based flow such as in [5][17]=-=[21]-=-, the motion for the pixel at in one frame to a successive frame is defined to be the displacement of the patch of pixels centered at , out of possible displacements (where is an arbitrary parameter d... |

10 |
Using flow field divergence for obstacle avoidance in visual navigation
- Nelson, Aloimonos
(Show Context)
Citation Context ...t system. 9 Discussion Some researchers [20][33][34] have proposed using divergence or flow derivatives for visual cues, but did provide real-time implementations of these ideas. Nelson and Aloimonos =-=[27]-=-[28] used directional flow divergence for stop-and-look obstacle avoidance (not real-time smooth driving). Their environments were simplified, and they did not demonstrate extensive robust behavior ov... |

8 | Ecological robotics
- Duchon, Warren, et al.
- 1998
(Show Context)
Citation Context ...ook obstacle avoidance (not real-time smooth driving). Their environments were simplified, and they did not demonstrate extensive robust behavior over extended periods of time. Duchon, et al. [14][15]=-=[16]-=- demonstrated flow and flow-derived time-to-contact for free wandering at 5 cm/s for as long as 5 minutes. In a second implementation, speeds of up to 30 cm/s have been achieved, and unrecorded trials... |

7 |
Two-dimensional root structures and convergence properties of the separable median filter
- Nodes, Gallagher
- 1983
(Show Context)
Citation Context ...sion of an object's full spatial or temporal extent could create the illusion of an open space and cause a collision. This corner preservation can also prevent distinct objects from blurring together =-=[29]-=-, which could be equally undesirable because some valid pathways may otherwise falsely appear to be blocked. When performing the separable temporal median filter, the order of the two 1-D median opera... |

7 |
Visual servoing from 2D image cues
- Raviv, Herman
- 1993
(Show Context)
Citation Context ...ormation is only represented in 2-D image coordinates. The control algorithms directly use observable image information represented in the 2-D image sequence; a 3-D reconstruction is not required [18]=-=[31]-=-. Such an approach is particularly useful in closing control loops with vision at lower levels of a multi-level control system. The limited body-centered data that are used are mapped into the local v... |

6 | Space-time trade-offs for adaptive realtime tracking
- Camus, Bülthoff
- 1991
(Show Context)
Citation Context ...evised 1999 6 A significant limitation with the traditional correlation-based algorithm is that its time complexity grows quadratically with the maximum possible displacement allowed for the pixel [5]=-=[9]-=- (Figure 3 (a)). Intuitively, as the speed of the object being tracked doubles, the time taken to search for its motion quadruples, because the area over which we have to search is equal to a circle c... |

6 |
A Parallel Algorithm for Real-Time
- Blthoff, Little, et al.
- 1989
(Show Context)
Citation Context ... point of our implementation. DRAFT, submitted 1997, revised 1999 4 (b) Robot’s view of the lab through its 115-degree wide angle lens Figure 2 Robot and Its ViewsIn correlation-based flow such as in =-=[5]-=-[17][21], the motion for the pixel at [,] xy in one frame to a successive frame is defined to be the displacement of the patch Pν of ν× ν pixels centered at [,] xy , out of ( 2η+ 1) × ( 2η + 1) possib... |

4 |
Parallel Dense Depth from Motion on the Image Understanding
- Weems
- 1993
(Show Context)
Citation Context ...ion. (a) Test mobile robot (b) Robot's view of the lab through its 115-degree wide angle lens Figure 2 Robot and Its View DRAFT, submitted 1997, revised 1999 5 In correlation-based flow such as in [5]=-=[17]-=-[21], the motion for the pixel at in one frame to a successive frame is defined to be the displacement of the patch of pixels centered at , out of possible displacements (where is an arbitrary paramet... |

3 |
New visual invariants for terrain navigation without 3-D reconstruction
- Young, Hong, et al.
- 1998
(Show Context)
Citation Context ...ted 1997, revised 1999 15 (3) where is the depth of the object in the environment relative to the camera, and and are the translational and rotational motion of the environment relative to the camera =-=[35]-=-. This camera frame is a right-handed coordinate system centered behind the image plane with the positive X direction extending leftward, positive Y upward, and positive Z forward. The divergence of a... |

3 | Parallel dense depth from motion on the image understanding architecture - Dutta, Weens - 1993 |

2 |
Minimalist vision for navigation
- Herman, Nashman, et al.
- 1997
(Show Context)
Citation Context ... information is only represented in 2-D image coordinates. The control algorithms directly use observable image information represented in the 2-D image sequence; a 3-D reconstruction is not required =-=[18]-=-[31]. Such an approach is particularly useful in closing control loops with vision at lower levels of a multi-level control system. The limited body-centered data that are used are mapped into the loc... |

2 |
Optic Flow," Vision Research
- Koenderink, Doorn
- 1986
(Show Context)
Citation Context ...ollision-free (however crude) mobility as a base of competence, it is promising enough to be considered as a low-level competence in a goal-directed mobile robot system. 9 Discussion Some researchers =-=[20]-=-[33][34] have proposed using divergence or flow derivatives for visual cues, but did provide real-time implementations of these ideas. Nelson and Aloimonos [27][28] used directional flow divergence fo... |

2 |
Neural Networks for Mobile Robot Navigation
- McCann
- 1995
(Show Context)
Citation Context ...y a factor of 4 on a side is performed on the host workstation itself. The above algorithm returns quantized optical flow values. Although this is adequate for various robotic vision tasks [6][14][15]=-=[25]-=-, it is not sufficient for our application, because the calculation of divergence requires the ability to measure spatial derivatives of the optical flow. Because quantized optical flow is basically a... |

1 |
Real-Time Quantized Optical Flow”, to appear in The Journal of Real-Time Imaging (special issue on Real-Time Motion Analysis
- Camus
- 1996
(Show Context)
Citation Context ...g system, it is important to have sufficient processor idle time available to buffer the overhead of the operating system, otherwise the image capture frame rate could vary from frame to frame. Camus =-=[6]-=- describes a robust, real-time correlation-based optical flow algorithm which returns dense data even in areas of low texture. This algorithm is the starting point of our new implementation. In correl... |