Human fixations detection model in video-compressed-domain based on MVE and OBDL
Yongjun Li, Xuemei Lei, Yong Liang, Jing Chen · 2018
This paper deals with the problem of compressed-domain human fixations detection in the video sequences, and presents a fast and efficient algorithm based on Motion Vector Entropy (MVE) and Operational Block Description Length (OBDL). The two features are obtainable from the compressed video bitstream with partial decoding, and generate the feature maps. The two feature maps are processed, and generate MVE map and OBDL map respectively. Then the processed maps are fused. In order to further improve the global saliency detection, the fused map is worked by the Gaussian model whose center is determined by the feature values. The validation and comparison are made by several accuracy metrics on two ground truth datasets. Experimental results show that the proposed saliency detection model obtains superior performances over several state-of-the-art compressed-domain and pixel-domain algorithms on evaluation metrics. Computationally, our algorithm achieves the real-time requirements of the saliency detection.