This paper describes the implementation of a realtime maximally stable extremal region (MSER) detector. In order to reach real-time performance, both algorithmic and memory issues have been addressed. The union-find algorithm, which is the heart of the MSER detector, is extended to create linked regions that significantly decrease the time to extract MSERs. Hash indexed memory structures are used to locate stored regions fast while keeping the amount of stored data low. The design is verified by including it in a demonstrator circuit. Timing and memory requirements are presented for the demonstrator and as a function of image resolution

In this paper, we discuss the design and implementation of a Field-Programmable Gate Array (FPGA) based stereo depth measurement system that is capable of handling a very large disparity range. The system performs rectification of the input video stream and a left-right consistency check to improve the accuracy of the results and generates subpixel disparities at 30 frames/second on 480 × 640 images. The system is based on the Local Weighted Phase- Correlation algorithm [9] which estimates disparity using a multi-scale and multi-orientation approach. Though FPGAs are ideal devices to exploit the inherent parallelism in many computer vision algorithms, they have a finite resource capacity which poses a challenge when adapting a system to deal with large image sizes or disparity ranges. In this work, we take advantage of the temporal information available in a video sequence to design a novel architecture for the correlation unit to achieve correlation over a large range while keeping the resource utilisation very low as compared to a naive approach of designing a correlation unit in hardware.

In this work, we propose a denoising scheme to restore images degraded by CCD noise. The CCD noise model, measured in the space of incident light values (light space), is a combination of signal-independent and signal-dependent noise terms. This model becomes more complex in image brightness space (normal camera output) due to the nonlinearity of the camera response function that transforms incoming data from light space to image space. We develop two adaptive restoration techniques, both accounting for this nonlinearity. One operates in light space, where the relationship between the incident light and light space values is linear, while the second method uses the transformed noise model to operate in image space. Both techniques apply multiple adaptive filters and merge their outputs to give the final restored image. Experimental results suggest that light space denoising is more efficient, since it enables the design of a simpler filter implementation. Results are given for real images with synthetic noise added, and for images with real noise.

We propose a denoising scheme to restore images degraded by CCD noise. Typically, restoration algorithms assume a linear mapping between the incident light space and image space. However, in practice, a camera response function performs a non-linear mapping on the sensor output and, as a result, the sensor noise model becomes more complex in the image space. We correct for non-linearity by mapping the corrupted image into "light space", where the relationship between the incident light and light space values is linear. To reduce the sensor noise we accurately model the CCD sensor noise by using the photon transfer curve. We then develop a combination of adaptive filters based on the estimated noise model in light space. Our adaptive system demonstrates efficient noise removal performance in uniform regions, while preserving edges and fine details.

This paper describes the implementation of a stereo depth measurement algorithm in hardware on field programmable gate arrays (FPGAs). This system generates 8 bit sub-pixel disparities on 256 by 360 pixel images at video rate (30 frames/sec). The algorithm implemented is a multi-resolution, multi-orientation phase-based technique called local weighted phase-correlation (Fleet, 1994). Hardware implementation speeds up the performance more than 300 times that of the same algorithm running in software. In this paper, we describe the programmable hardware platform, the base stereo vision algorithm and the design of the hardware. We include various trade-offs required to make the hardware small enough to fit on our system and fast enough to work at video rate. We also show sample outputs from the functioning hardware. Although this paper is specifically focused on phase-based stereo vision FPGA realizations, most of the design issues are common to other DSP and vision applications.

A new technique for fast pattern recognition using normalized grey-scale correlation (NGC) is described. While NGC has traditionally been slow due to computational intensity issues, the introduction of both a pyramid structure and a local estimate of the correlation surface gradient allows for recognition in 10-50 ms using modest microcomputer hardware. The algorithm is designed to analyze the target off-line prior to starting the search. Issues surrounding determining an appropriate depth for the pyramid representation and performing sub-pixel localization of the target instance are discussed. The speed and robustness of the method makes it attractive for industrial applications.

Given estimates of the motion field (optic flow) from an image sequence, it is possible to recover translational direction, T/spl I.oarr/, using a variety of techniques. One such technique, known as "subspace methods, generates constraints which are perpendicular to T/spl I.oarr/, so that two distinct constraints allow a solution for T/spl I.oarr/. In practice many constraints are used in a least-squares solution, but it has been observed that the recovered estimates for T/spl I.oarr/ are biased towards the optical axis. While the cause of the bias is well known, previous attempts to remove it have been flawed. This paper outlines a new method which removes the bias. The technique is simple to apply and computationally efficient.

This paper examines the use of the EM algorithm to perform motion segmentation on image sequences that contain independent object motion. The input data are linear constraints on 3-D translational motion and bilinear constraints on 3-D translation and rotation, derived from computed optical flow using subspace methods. The problems of outlier detection, deciding how many processes, and the initial guesses for the EM algorithm are considered. Results obtained from an image sequence are presented. 1 Introduction In order for an observer to navigate in its environment, it is important that the observer can detect other independently moving objects and avoid collisions. The motion of the observer complicates this task. For the purpose of this paper we divide image motion into two categories: egomotion and motion due to independent moving objects. Egomotion is defined as the image motion induced by an observer moving through a static environment. Motion due to independently moving objects ...