基於光流計算之多目標物追蹤

Abstract

在電腦視覺的應用上，偵測運動物體為重要的研究課題之一，本論文探討如何從一連串的影像（視訊）中擷取運動物體的資訊，此一研究成果可應用於機器人導航、目標追蹤、影像編碼等方面。 在本論文中，我們提出了一種抗雜訊的光流估測方法。此方法先分析在速度向量場的光流限制式的直線交點分佈，去除誤差較大的光流限制式，再利用最小平方差方法求出光流估測值。這些估測值可被用來判斷是移動物體或是背景。當背景為平滑移動或移動物體發生遮蔽的情形時，本方法仍可有效追蹤影像中的移動物體。此方法利用影像金字塔處理，以減少計算量和雜訊的影響。 從實驗結果可清楚發現本論文提出的方法得到較正確的光流估測。本論文中做了多個真實影像的實驗，經由實驗結果知道，若物體呈像面積不要太小而且形變不算太大下，所提的方法可以有效的追蹤影片中之移動目標物體。

Detecting moving objects is one of the important research topics in computer vision. This thesis investigates how to obtain the information of moving objects from an image sequence. The research results can be applied to robot navigation, target tracking and image coding etc. In this thesis a robust against noise method of the optical flow estimation is shown. The technique is based on the optical flow constraint, it analyses the intersection of optical flow constraint on velocity field then discarding the erroneous optical flow constraint. Appling the least squares method solves the constraint line equations. The optical flow estimation used to separate the target from the background. When the background moving smoothly or the moving objects occlude, the proposed algorithm is also tracking successful. In order to decrease the amount of calculation and the influence of noise, using the image pyramid pre-process. From the experimental results, on both synthetic and real sequences, show clearly the proposed approach performs quite better on the accuracy of optical flow estimation. Several real videos are experimented: Under the assumption of the size and deformation of targets, proposed approach efficient tracking multi-target.