This paper presents a detailed study of four image descriptors (SURF, SIFT, BRISK and FREAK) in the context of real-time detection of abandoned objects using a moving camera. In this scenario, captured frames are compared to a reference video, and noticeable differences among the two videos are associated to an abandoned object. The image descriptors allow a simple and robust image representation, retaining most relevant features for proper registration and alignment, enabling a comparison between two video frames corresponding to the same scene. Performances of these fours schemes are assessed in terms of processing time and detection efficiency considering the OpenCV implementations of the methods. A modification is also considered for the image descriptors, which restricts the correspondences between the two images to the same representation scale. Experiments on three pairs of videos show the improvements achieved by the proposed system configuration.