Commentary Paper 2 on "Robust Unattended and Stolen Object Detection by Fusing Simple Algorithms"

The technique discussed in this article proposes to distinguish between unattended and stolen objects by combining shape and appearance similarity measures of foreground objects observed in consecutive frames of a video. Static objects are detected by examining trajectories and people are removed from consideration. Object shape boundary is first refined using active contours. Shape similarity is then defined by computing gradient magnitudes along the object boundaries and counting how many boundary pixels have values higher/lower than predefined thresholds. Appearance similarity is based on differences between histograms using the foreground mask on the current and background images. Probabilities are defined assuming the shape and appearance measures follow the Gaussian distribution (with trained parameters). Final measures of unattended/stolen objects are produced by averaging the probabilities and used to classify static-nonhuman objects. Experiments show that combining the three measures gives better results than using each of the measures alone.