Real time contour tracking with a new edge detector

Abstract

In this paper, a new system for real time contour tracking is presented. If a rough contour of the desired structure is available on the first image of a sequence, the system can automatically outline the contours on the subsequent images at video rate. The method we used is based on a new edge detector which was obtained by the generalization of the first order absolute central moment operator. The new algorithm proved to be very robust to noise and fast enough to be implemented in real time. The contour tracking procedure was implemented on an integrated software/hardware platform composed of a personal computer equipped with a digital signal processing board. The system can capture an analog video signal with a resolution of 512×512 pixels, 25 frames/s, process the data and display the results in real time. A graphical user interface is also available to interact with the system. Tests on images of the descending thoracic aorta and of a carotid, recorded by echocardiography, are reported. The cross-sectional area of the aorta and the diameter of the carotid were computed in real time and plotted on the user interface. The system proved to be a useful tool for the investigation of vascular mechanisms.

Introduction

Contour tracking on sequences of images is much debated in image analysis. The topic is of particular relevance in various application areas, such as robotics, video surveillance, video conferencing, human–computer interaction and medical imaging [1], [2], [3], [4], [5], [6], [7]. In general, the goal is to track the silhouette of a moving object throughout a sequence of images. Several papers exploit a contour tracking technique to approach the problem of object tracking, focusing their attention on the position of the object rather than on its shape. People and car tracking systems [1], [2], whose purpose is to locate the position of one or more moving objects inside a scene, are typical examples. Such papers are usually not preoccupied with the exact detection of the contour. On the contrary, other applications focus on an accurate definition of the shape of the object under investigation. Examples can be found in medical imaging, where the silhouette of an anatomical structure provides information of clinical interest [3], [4], [5]. Is not an easy task to track the contour of an object automatically. The complexity of the problem remarkably increases when the object is not rigid, the movement is three-dimensional (3D), the background of the image is complex or the view of the object is partially occluded [6], [7]. Further difficulties arise when real time performances are required.

Despite the large amount of literature, a general solution to the problem of contour tracking does not exist, so the issue is still an open problem. In recent years, methods based on the active contours have often been used to track the contours of structures under investigation in image sequences [8]. These methods usually require the minimization of an energy function associated to a deformable continuous curve. In general, two energy terms are considered: (i) an internal energy, which increases when the deformation of the curve increases and (ii) an external energy, which increases when the distance between the curve and the contour points increases. A starting curve which approximates the contour is given and the final curve is a compromise between contour smoothness and closeness to the data. As regards the external energy, the magnitude of the gradient of Gaussian (GoG) is commonly used to weigh the distance between the curve and the contour points. One of the advantages of using an approach based on the active contours is the reduction of the amount of data that must be processed, which is limited to a neighborhood of the starting curve. Nevertheless, when real time performances are required, the complexity of the algorithms must be further reduced. In [9], [10] the task of energy minimization on a two-dimensional (2D) plane was reduced to a mono-dimensional (1D) search problem. Such an approach dramatically reduces the computational complexity, thus allowing a real time execution of the procedure on a standard platform such as a personal computer. However, the robustness to noise usually decreases also since in this case the regularizing low pass filter is a 1D filter and, when the aperture of the filter is chosen, the number of pixels involved in the filtering process decreases with respect to a 2D filtering process.

In this paper, a new method for real time contour tracking is presented. The approach is similar to those based on active contours since this method also finds the “true” contour starting from an approximate contour and includes a regularization procedure. However, both the edge detector and the associated localization strategy, which we used to locate and track the contour through the image sequence, are completely new. The research of a new edge detector started a few years ago with the study of a generalization of the absolute central moment operator [11], [12]. With respect to the standard edge detectors based on the GoG or on the Laplacian of Gaussian (LoG), the new operator provided better results at image key points such as lines, corners, junctions and isolated points. When focusing our attention on contour tracking in real time an additional property of the absolute central moment was found: when a starting point close to a gray-level discontinuity is given, the nearest point of the discontinuity can be localized with an iterative and very fast algorithm. Section 2 describes the mathematical operator and the contour tracking procedure, while Section 3 reports a comparison with the GoG.

In addition to the problem of developing an efficient algorithm, real time contour tracking also involves aspects of hardware/software design. This is a typical aspect of real time imaging applications, where the processing of a large amount of data in a short space of time is required [13]. It is easy to understand that the main issue is the availability of enough processing power to ensure the execution of the algorithms within a given time. Recently, several authors faced the problem by using standard workstations, as the modern general-purpose processors achieve remarkable performances in terms of computational speed. The advantage of such solutions is that they use a well-known general-purpose platform, which is easily programmable by means of powerful software development tools. However, when higher performances are required, special purpose architectures must be adopted, which can provide far more processing power than a standard workstation. The solutions most commonly used are based on programmable devices, such as DSPs or FPGAs, by means of which powerful and quite flexible hardware platforms can be obtained. The main drawback in using such platforms is that implementing algorithms may be difficult, as a specific knowledge of the hardware in use is necessary. Besides the processing power, there are other aspects that must be taken into account in the design of the system. In real applications, the images are often available as an analog video signal, which must be captured, processed and then converted again into an analog signal for the display on a monitor. Simultaneously capturing, processing and displaying the data in real time is a task that requires the ability to quickly manage a large amount of data. The capability of moving the data and synchronizing the different activities is a complex job and requires the use of suitable hardware and software tools. This problem, which is neglected in some papers, can cause some difficulties in standard platforms that are not designed to perform real time tasks.

In other words, the choice of the hardware/software platform greatly depends on the specific requirements of the application. In this paper we adopted a special hardware platform, which was obtained simply by integrating a standard personal computer with a video processing board. Implementing algorithms on a custom hardware required a larger effort with respect to the use of a standard workstation. However, the processing power of the platform allowed us to obtain real time execution without reducing both the quality of the images, which are 512×512 pixels and 25 frames/s, and the size of the mathematical operator, which is a robust-to-noise 2D operator. Section 4 describes both the hardware platform and the software implementation.

The system was largely tested on sequences of ultrasound images. Results are presented on the contour tracking of aorta cross-sections and of carotid longitudinal sections. Furthermore, tests are presented on the tracking of a moving object.