A fast and efficient saliency detection model in video compressed-domain for human fixations prediction

Research and application of human fixations detection in video compressed-domain have gained an increasing attention in the latest years. However, both prediction accuracy and computational complexity still remain a challenge. This paper addresses the problem of compressed-domain video human fixations prediction based on saliency detection, and presents a fast and efficient algorithm based on Residual DCT Coefficients Norm (RDCN feature) and Operational Block Description Length (OBDL feature). These two features are directly extracted from the compressed bit-stream with partial decoding, and are normalized. After spatial and temporal filtering, the normalized salient maps are fused by the dynamic fusion coefficients with variation of quantization parameters. Then the fused salient map is worked by Gaussian model whose center is determined by the feature values. The proposed saliency detection model for human fixations prediction combines the accuracy of the pixel-domain saliency detections with the computational efficiency of their compressed-domain counterparts. The validation and comparison are made by several accuracy metrics on two ground truth datasets. Experimental results show that the proposed saliency detection model for human fixations prediction obtains superior performances over several state-of-the-art compressed-domain and pixel-domain algorithms on evaluation metrics. Computationally, our algorithm achieves a speed-up of over 10 times as compared to similar algorithms, which illustrates it appropriate for in-camera saliency estimation.