應用單像機序列影像於物件定位與追蹤

Abstract

隨著科技的進步，影像技術的應用除了更普及之外也更加日新月異，舉凡遙測影像的變遷偵測、無人車的障礙物偵測和監視攝影機的行人偵測等皆是影像用於偵測之應用。然而像機拍攝時可能因為自然或人為因素而發生運動。此外，拍攝場景內可能發生光影變化，進而影響後續的偵測成果。為了要能在像機發生運動的情況下使用序列影像進行移動物體偵測，本研究透過解算多張影像間的相對方位，以改正像機運動對於移動物體偵測的影響。另外，為了避免因為場景內光影變化造成偵測錯誤的發生，本研究發展基於HSI色彩空間與背景相減法之移動物體偵測方法。當取得移動物體於影像上的位置，可透過影像資料所提供的資訊搭配共線條件進行三維物體追蹤。藉由假設物體是在一個高度未發生變化的地面上運動，移動物體的位置與高度皆可根據所提方法取得。實驗結果顯示，根據所提方法只需要單張影像即可完成三維物件追蹤，且精度可達公分級。本研究方法能降低當前移動物體偵測的限制，並擴展單像機的使用於更廣泛的應用。

With the rapid development of technology, the image processing technique is changing frequently and is becoming more common nowadays. There are numerous applications of this technique, such as change detection that use remote sensing image, obstacle detecting system of unmanned vehicle, and moving people detection with surveillance cameras. However, the view of camera is often changed due to natural factors or human activities. In addition, inconsistency of images with the same view may also happen as result of luminance variation. In order to correctly detect the moving objects from image sequence captured by a moving camera, the moving correction of camera has to be done firstly. This can be accomplished by acquiring the relative orientation between two image frames. Next, to avoid the detection error caused by luminance variations or shadow effect, a moving object detection approach based on the HSI color space and background subtraction algorithm is developed in this study. Once the 2D position of moving objects is available, 3D object tracking can be performed by using the information from image data and the collinearity condition. By assuming that the object is moving on the ground without changing in altitude, the location as well as the attitude of the moving objects can be obtained based on the proposed approach. The experiment results indicated that 3D object tracking of centimeter-level accuracy can be achieved with the proposed approach, in which only single image is required. Consequently, this approach is capable of reducing the limitation of current moving object detection technique and extending the usage of a single camera for further applications.