Detecting sudden moving objects in a series of digital images with different exposure times

摘要

This paper presents an algorithm to detect sudden objects appearing within a set of digital images obtained at different exposures to create a high dynamic range (HDR) image. While some previous work has focused on detecting moving objects within a scene, the majority cannot handle exposure variability. The few techniques developed specifically for HDR images track moving objects that have smaller movements within a scene compared to an abrupt object that appears and disappears quickly. Further, the algorithm advances existing methods because it does not require: 1) robust estimation of a camera response function, 2) supervision of objects in the scene such as explicit object detection and tracking, and 3) selection of a reference image. In this approach, every image in the set is first partitioned into equal size patches. Next, the properties (e.g., histograms of oriented gradients, HOG) of values within the window of the same patch are compared between the images to identify differences. Finally, a statistical classifier is developed to recognize significant differences between patch descriptors and identify patches containing sudden objects. This statistical classifier makes it possible to define confidence levels for categorizing patches into a moving object or not. A sensitivity analysis indicated that the best performance occurs when using four or six digital images. However, the optimal patch size is dependent on the size of the moving object to be detected. Hence, a mechanism is introduced to estimate the range of reasonable patch sizes given an image.