Saliency Analysis for Car Detection Based on 2-D Entropy and Velocity Prior

In this paper, an efficient saliency analysis method based on entropy theory and wavelet analysis is proposed, which can be used for car detection in traffic video. In the authors' method, the wavelet theory is used to detect the global saliency parts on a single frame in the video. The entropy theory is applied to choose the best saliency map among three color channels. Besides, the use of velocity information can enhance the objects close to observers, which is consistent with humans’ visual habit. Experimental results show that the authors' method can achieve excellent results in terms of receiver operating characteristic (ROC) curve, the area under the curve (AUC) score, linear correlation coefficient (CC) score, and normalized scanpath saliency (NSS) score, as compared to other state-of-the-art methods.