Information theoretic approach for low-complexity adaptive motion estimation

Due to the high demand of wireless video applications, the video compression by mobile devices with limited energy supply, low processing capability and narrow channel bandwidth has become an important issue. The constraints of these mobile devices demand video compression systems with higher efficiency and lower energy consumption. Furthermore; in the emerging wireless video sensor networks, sensor nodes are battery-operated and responsible for data collection: these sensor nodes have far more stringent power constraints than the data sink nodes. Therefore, energy conservation of sensor nodes is crucial for the longevity of wireless video sensor networks. It motivates the development of video compression schemes of extremely low computation complexity to be used in video sensor nodes.; Motion estimation is a key component in a video encoder. By predicting the subsequent frames from reference frames, motion estimation squeezes temporal redundancy in the video sequence and achieves compression efficiency. More importantly, motion estimation is a major concern for designing energy-efficient video encoders since in most cases motion estimation constitutes roughly 70% of the computational load on a video encoder[4]. Hence, motion estimation becomes the central issue of this dissertation.; In this dissertation, we propose a new technique for motion estimation with the objective of achieving low computation complexity. The proposed technique is motivated by recent research on information theoretic learning where Renyi's definition of entropy and mutual information are used as optimization criteria. In our methods, the motion estimation problem is modeled as an optimization problem under an adaptive system framework with motion vectors as the parameters of the systems to be optimized. This adaptive system architecture allows the information obtained from the processing of the previous frames to be stored and used to aid the future estimation effort. Also, we propose to use two newly developed optimization criteria for this purpose. In this dissertation, we present our work within a unified framework for adaptive motion estimation with information theoretical learning criteria, which leads to various motion estimation schemes. As a prominent feature of the proposed schemes, the computational complexity is reduced significantly compared to the traditional methods, making our motion estimation algorithms suitable for video applications under stringent resource constraints such as wireless video sensor networks.