Segmentation and Tracking of Moving Objects in a Image Sequence

研究了一种光流矢量检测算法，该算法通过对滤波后的图像应用多个滤波器并求解所导出的方程来获得每个像素处的光流矢量。一个专用设备的开发，用于执行该算法在一个高速。该装置由两个DSP板串联而成，通过流水线处理，每秒15次提取图像中的流矢量。编制了DSP程序，并通过实验确定了最佳滤波器。提出了一种利用多个滤波器提取可靠的流矢量并获得匀速运动区域的方法。如果两幅图像中的物体边界不明确，则通过使用两幅以上的图像来确定准确的区域和运动。对真实的图像进行运动目标跟踪处理。当指定对象区域时，从光流获得对象的质心，并且通过控制相机的摇摄和倾斜来跟踪对象，使得对象可以在图像的中心。该算法在专用设备中实现，并对运动目标进行了真实的实时跟踪。为了处理长序列图像中的运动变化，研究了基于最小描述长度（MDL）原则的时空分割方法。在该方法中，较长的边缘首先被提取，他们被分割的空间和时间。然后利用运动模型，基于MDL原则进行最终的时空分割。

An algorithm for optical flow vector detection is studied which obtains the flow vector at each pixel by applying multiple filters and solving the equations derived from the filtered images. A special purpose device is developed for performing the algorithm at a high speed. The device consists of serially connected boards with two DSPs, and extracts flow vectors in the image 15 times a second by pipeline processing. The DSP program is made and the optimal filters were determined by experiments. A method is developed for extracting reliable flow vectors with multiple filters and obtaining regions with the uniform motion. If boundaries of objects are ambiguous from two images, accurate regions and the motions are determined by using more than two images. Moving object tracking processing is applied to real images. When an object region is specified, the centroid of the object is obtained from the optical flow, and the object is tracked by controlling the pan and tilt of a camera so that the object may be in the center of the image. The algorithm is implemented in the special purpose device, and a moving object is tracked in real time. In order to deal with motion changes during a long sequence of images, spatio-temporal segmentation is studied based on the Minimum Description Length (MDL) principle. In the method, longer edges are first extracted, and they are segmented spatially and temporally. Then using models of the movement, the final spatio-temporal segmentation is performed based on the MDL principle.