Motion saliency detection using a temporal fourier transform

• We propose a new Temporal Fourier Transform based motion saliency detection method.
• Variations in the phase spectrum correspond to the time-varied “pop-out”.
• Variations in the phase spectrum are identified to obtain the salient location.
• We achieve a better tradeoff between the accuracy and the computational cost.
• Our model is independent of features, background models, or prior knowledge.

Motion saliency detection aims at detecting the dynamic semantic regions in a video sequence. It is very important for many vision tasks. This paper proposes a new type of motion saliency detection method, Temporal Fourier Transform, for fast motion saliency detection. Different from conventional motion saliency detection methods that use complex mathematical models or features, variations in the phase spectrum of consecutive frames are identified and extracted as the key to obtaining the location of salient motion. As all the calculation is made on the temporal frequency spectrum, our model is independent of features, background models, or other forms of prior knowledge about scenes. The benefits of the proposed approach are evaluated for various videos where the number of moving objects, illumination, and background are all different. Compared with some the state of the art methods, our method achieves both good accuracy and fast computation.