Experimental Study on Realtime Object Tracking Stystem

(1)快速可靠地提取光流在每个位置，由连续输入图像与一组方向选择空间滤波器进行卷积得到的多个约束方程计算出一个流向量及其不确定性测度。(2)基于光流的目标跟踪：通过更新一个矩形窗口来包围具有相似流矢量的目标区域，从而实现对目标的跟踪。在多目标跟踪中，为每个目标分配一个跟踪窗口。当两个对象重叠时，通过比较重叠区域内的流量与每个窗口内的流量来确定前景对象。我们还开发了一种通过整合空间和时间信息来跟踪复杂背景中的物体的方法。(3)基于光流和边缘的目标跟踪根据光流得到的物体运动边界，在物体周围搜索物体的边缘。通过对提取的边缘进行累加，增量得到目标的整体轮廓。当两个运动相似的物体重叠时，通过比较预测轮廓和实际轮廓来确定前景物体。(4)基于光流和视差的目标跟踪视差被计算为一对立体图像之间的光流。利用光流、视差和预测的目标位置，通过贝叶斯推理提取目标区域。即使单独使用流矢量或不均匀跟踪可能失败，该方法也可以跟踪目标。(5)实时跟踪系统的开发实验跟踪系统由基于dsp的实时图像处理器和有源立体摄像机头组成。图像处理器包含多个并行工作的DSP板，可以在接近一帧速率的情况下处理复杂的图像。利用该系统实现了对目标的实时跟踪。

(1) Fast and Reliable Extraction of Optical FlowAt every location, a flow vactor and its uncertainy measure are calculated from multiple constraint equations, which are obtained by convolving consecutive input images with a set of orientation-selective spatial filters.(2) Object Tracking Based on Optical FlowAn object is tracked by updating a rectangular window circumscribing the object region which has similar flow vectors. In multiple object tracking, a tracking window is assigned to each object. When two objects overlap, the foreground object is determined by comparing the flow in the overlapping region with that in each window. We also developed a method to track an object in a complicated background by integrating information over space and time.(3) Object Tracking Based on Optical Flow and EdgesEdges of an object are searched for around the object's motion boundary obtained from optical flow. The whole contour of the object is incrementally obtained by accumulating extracted edges. When two objects with similar motion overlap, the foreground object is determined by comparing the predicted contrours and the actual controur.(4) Object Tracking Based on Optical Flow and DisparityDisparity is calculated as optical flow between a pair of stereo images. An object region is extracted by Baysian inference using optical flow, disparity, and the predicted target location. The method can track an object even if the tracking with either flow vectors or disparities alone may fail.(5) Development of Realtime Tracking SystemAn experimental tracking system is composed of a DSP-based realtime image processor and an active stereo camera head. The image processor, which contains multiple DSP boards working in parallel, enables complex image processing in nearly a frame rate. We succeeded in realtime object tracking using the system.