Theodore(Ted) Camus

Accuracy vs Efficiency Trade-offs in Optical Flow Algorithms

There have been two thrusts in the development of optical flow algorithms. One has emphasized higher accuracy; the other faster implementation. These two thrusts, however, have been independently pursued, without addressing the accuracy vs efficiency trade-offs. Although the accuracy?efficiency characteristic is algorithm dependent, an understanding of a general pattern is crucial in evaluating an algorithm as far as real-world tasks are concerned, which often pose various performance requirements. This paper addresses many implementation issues that have often been neglected in previous research, including temporal filtering of the output stream, algorithms flexibility, and robustness to noise, subsampling, etc. Their impacts on accuracy and/or efficiency are emphasized. We present a survey of different approaches toward the goal of higher performance and present experimental studies on accuracy vs efficiency trade-offs. A detailed analysis of how this trade-off affects algorithm design is manifested in a case study involving two state-of-the-art optical flow algorithms: a gradient and a correlation-based method. The goal of this paper is to bridge the gap between the accuracy- and the efficiency-oriented approaches.

Real-Time Quantized Optical Flow

Algorithms based on the correlation of image patches can be robust in practice but are computationally intensive due to the computational complexity of their search-based nature. Performing the search over time instead of over space is linear in nature, rather than quadratic, and results in a very efficient algorithm. This, combined with implementations which are highly efficient on standard computing hardware, yields performance of 9 frame/sec on a scientific workstation. Although the resulting velocities are quantized with resulting quantization error, they have been shown to be sufficiently accurate for many robotic vision tasks such as time-to-collision and robotic navigation. Thus, this algorithm is highly suitable for real-time robotic vision research.

Real-time single-workstation obstacle avoidance using only wide-field flow divergence

A real-time robot vision system is described 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. The entire system is implemented on a single ordinary UNIX workstation without the benefit of real-time operating system support. Dense optical flow data are calculated in real-time across the entire wide-angle image. The divergence of this optical flow field is calculated everywhere and used to control steering and collision behavior. Divergence alone has proven sufficient for steering past objects and detecting imminent collision. The major contribution is the demonstration of a simple, robust, minimal system that uses flow-derived measures to control steering and speed to avoid collision in real time for extended periods. Such a system can be embedded in a general, multi-level perception/control system.

Real-time quantized optical flow

Algorithms based on the correlation of image patches can be robust in practice but are computationally intensive due to the computational complexity of their search-based nature. Performing the search over time instead of over space is linear in nature, rather than quadratic, and results in a very efficient algorithm. This, combined with implementations which are highly efficient on standard computing hardware, yields performance of over 5 frames per second on a scientific workstation. Although the resulting velocities are quantized with resulting quantization error, they have been shown to be sufficiently accurate for many robotic vision tasks such as time-to-collision and robotic navigation. Thus, this algorithm is highly suitable for real-time robotic vision research.

Calculating time-to-collision with real-time optical flow

Currently two major limitations to applying computer vision to real-time robotic vision tasks are robustness in unsimulated and uncontrolled environments, and the computational resources required for real-time operation. In particular, many current visual motion detection algorithms (optical flow) are not suited for practical applications such as crash detection because they either require highly specialized hardware or up to several minutes on a scientific workstation. A recent optical flow algorithm [C94] however has been shown to run in real-time on a standard scientific workstation and yields very accurate time-to-contact calculations.

A real-time computer vision platform for mobile robot applications

Abstract A portable platform is described that supports real-time computer vision applications for mobile robots. This platform includes conventional processors, an image processing front-end system, and a controller for a pan/tilt/vergence head. The platform is ruggedized to withstand vibration during off-road driving. The platform has successfully supported experiments in video stabilization and detection of moving objects for outdoor surveillance, gradient-based and correlation-based image flow estimators, and indoor mobility using divergence of flow. These applications have been able to run at rates ranging from 3 to 15 Hz for image sizes from 64 × 64 to 256 × 256.

Space-time tradeoffs for adaptive real-time tracking.

Many current optical flow algorithms are not suited for practical implementations such as tracking because they either require massively parallel supercomputers, specialized hardware, or up to several hours on a scientific workstation. One particular reason for this is the quadratic nature of the search algorithms used in these problems. We present two modifications to these types of algorithms which can convert quadratic-time optical flow algorithms into linear-time ones. The first uses a variable image sampling rate which trades space for time and yields an algorithm that is at worst linear, and at best constant, in the speed of the moving objects in the image. This technique finds the fastest motion in an image and is ideal for tracking, since the fastest moving objects in a robots environment are generally the most interesting. The second modification extends this approach to create a multiple-speed optical flow field by transforming quadratic searches over space into linear searches in time. This space-time inversion has the effect of searching for faster moving objects in each image earlier than for slower moving ones, with additional effort being exerted to search for slower objects only when desired. A system of velocity masking allows a tradeoff of angular resolution (but not magnitude resolution) for an optical flow algorithm only linear in the range of velocities present.